Where the Money Went: The Biggest Technology Investments of the New Era

When people ask where the money has gone in the artificial intelligence boom, they usually expect a single answer, a company or a number. The honest answer is a stack. Money has gone into chips, and into the memory and packaging that make chips usable, and into the factories that make both. It has gone into buildings, and into the electricity and cooling that keep those buildings alive. It has gone into the bonds, leases and private loans that pay for all of it. It has gone into the laboratories that train models, and into the thousands of companies that build products on top of them. Each layer has a different business, a different risk and a different answer to the question of who is earning a return. This study walks down the stack one layer at a time as it stood at the end of June 2026. It uses company filings where they exist, such as Nvidia's quarterly reports, named analyses from banks and research bodies for the larger estimates, and trade reporting for the rest, and it separates what is reported from what is projected from what I think. I have tried to be careful with every figure, because this is a field where numbers are repeated until they harden into facts, and where the definitions behind them often differ from one source to the next.

Before descending, a framing idea. Every large investment cycle in history has had a structure in which the early money went into the physical bottleneck, the middle money went into financing the build, and the late money went into the applications that were supposed to justify it. Railways, electrification, fibre optic cable and housing all followed that sequence. The question that matters for each is not whether the technology works, since it usually does, but whether the revenue arrives before the financing runs out, and which layer holds the scarce resource when it does. Keep that in mind, because the answers in 2026 differ sharply by layer. The people earning extraordinary returns today are not the people who will necessarily earn them in five years, and the people bearing the most risk today are not always the ones who appear in the headlines.

Begin at the bottom, with the chips, because that is where the clearest returns are visible. Nvidia is the company at the centre of the cycle, and its filings give us what most of the rest of the stack lacks: audited numbers. For the fiscal year that ended in January 2026, it reported revenue of about 215.9 billion dollars, up 65 percent from the year before. In the first quarter of its new fiscal year, which ended in late April, it reported record revenue of 81.6 billion dollars, up 85 percent from a year earlier, with data centre revenue of 75.2 billion dollars, up 92 percent. Its gross margin was about 75 percent. Data centre networking revenue, which covers the equipment that connects thousands of chips so they behave like one machine, reached 14.8 billion dollars, nearly triple the level of a year earlier, and the company guided to about 91 billion dollars of revenue for the following quarter. It assumed no data centre compute revenue from China in that outlook, and reported that shipments of an older generation of data centre chips to China were zero in the quarter, compared with 4.6 billion dollars a year earlier.

Two details in those filings seem to me more revealing than the headline growth. The first is customer concentration and its change. The company said that the large cloud providers accounted for about half of data centre revenue, with the other half coming from a widening group of buyers: specialist AI cloud providers, industrial and enterprise customers, and sovereign programmes run by governments. The second is the margin. A hardware company earning a 75 percent gross margin at that scale is something with few precedents, and it tells you who holds the scarce resource. For now, that is the chip designer and the handful of suppliers behind it, and the pricing power it commands is the reason so much of the investment cycle's cash flows upward to one place. It also tells you where the risk to the cycle lies: if customers decide the price is too high, or build their own chips, or find they have more capacity than demand, this is the margin that gets compressed first.

The big buyers are, in fact, building their own chips and negotiating hard, and the chip layer is not as monolithic as one company's revenue suggests. Large cloud companies design custom accelerators for their own workloads, other chip makers have won large supply agreements, and a set of well funded start-ups is chasing specialised designs. One reported deal gave a large model developer the right to buy a sizeable stake in a rival chip designer at a nominal price in exchange for a commitment to purchase many gigawatts of its processors, which shows how unusual the financial engineering around chips has become. I mention these arrangements because they illustrate a feature of this cycle that I return to later: some of the money flowing in a circle between suppliers and their customers, with suppliers investing in the companies that buy their products. That is not improper, but it makes it harder to see how much demand is genuinely independent.

Step down a level and you reach the factories and the components around the chip, where the physical bottlenecks have been most severe. A modern AI accelerator is not a single piece of silicon. It is a processor combined with several stacks of high bandwidth memory on a shared base, using an advanced packaging technique that only a few facilities in the world can perform at scale. Analysts who follow the supply chain have described advanced packaging capacity, not wafer production, as the binding constraint on how many accelerators can ship. One widely read analysis estimated that the leading foundry's capacity for this packaging was roughly 35,000 wafers a month at the end of 2024 and was targeted to reach about 130,000 by the end of 2026, a quadrupling in two years that has little precedent in the industry, and reported that Nvidia had reserved a very large share of that capacity for 2026. I would treat the exact share as an estimate, since it comes from analysts and not from the companies, but the pattern is consistent across sources: capacity is sold out well in advance.

Memory is the second constraint at this level. The high bandwidth memory used in AI accelerators is made by a small number of manufacturers, and supply analyses published in the first half of the year describe the current generation as effectively sold out for 2026, with contract prices up by double digit percentages. The leading foundry itself announced price increases for its most advanced manufacturing processes beginning in 2026, and in April it guided to capital spending of 52 to 56 billion dollars for the year, with most of it directed at advanced process technology and a smaller share at packaging and testing. It has also committed to large additional investments in new factories in Arizona. What this means in practice is that the cost of every accelerator includes rising prices from the foundry, the memory makers and the packaging providers, and that each of those suppliers is earning healthy returns at the moment because the thing they sell is scarce. The returns in this layer are real, but they depend on scarcity, and scarcity is exactly what large capital spending is designed to end.

Now consider the plans of the companies buying all this equipment, because their spending sets the demand for everything below. Published estimates for 2026 capital spending by the five largest spenders, Microsoft, Alphabet, Amazon, Meta and Oracle, have been revised upward repeatedly. Early in the year, industry analysts put the combined figure at roughly 660 to 690 billion dollars, nearly double the 2025 level, with Amazon guiding to about 200 billion dollars, Alphabet to 175 to 185 billion, Meta to 115 to 135 billion and Microsoft to more than 120 billion. Estimates made after the companies' first quarter results in late April moved higher still, with some analysts putting the total close to or above 750 billion dollars. Goldman Sachs, in an earlier analysis, noted that spending had consistently exceeded forecasts and that annual capital spending of about 700 billion dollars would roughly match the peak intensity of the telecommunications build out of the late nineteen nineties, while a later Goldman model sketched a path in which annual spending on computing, data centres and power rises to roughly 765 billion dollars in 2026 and about 1.6 trillion dollars by 2031, totalling around 7.6 trillion dollars over those years. The same analysis noted that the cost of a megawatt of data centre capacity, which has historically been around ten million dollars, is moving toward fifteen to twenty million dollars for the newest facilities.

I would offer three observations about these numbers. The first is that they are plans, not outcomes, and plans in this cycle have usually been revised upward, which is both a sign of genuine demand and a sign that forecasters have been trailing events. The second is that the totals mix categories. Capital spending includes buildings, power equipment, networking and land as well as chips, so the share that ends up as revenue for chip companies is only part of the figure. The third is that the numbers are large relative to the cash flows of the companies spending them. Analysts have noted that spending by some of these companies is approaching the entire revenue of their cloud businesses, which means that returns will have to come from growth in revenue that has not yet materialised. A company with an enormous existing business can afford to wait, but the wait has a cost, and I will come to who pays it when we reach the financing layer.

So much for the first half of the stack, the layers that make and buy the equipment. The second half concerns the places where the equipment lives, the power that runs it, the money that pays for it, and the models and products that sit on top.

Begin with the buildings and the electricity, because this is the layer where money is hardest to turn into capacity quickly. A data centre can be built in roughly one to two years, but the grid connection that feeds it often takes three to seven. An industry survey by a fuel cell company, drawn from developers and utilities, found that utilities expected the time needed to deliver power to new sites to be one and a half to two years longer on average than developers assumed, and that onsite power generation was becoming a standard part of data centre planning, with roughly one in five campuses expected to exceed a gigawatt in scale by 2030. A grid engineering firm described lead times of ninety to one hundred and thirty weeks for the transformers and breakers that any connection requires, and a regulatory filing by a power equipment company, citing the International Energy Agency, noted that data centre construction takes two to three years while the generation and grid infrastructure to support it typically takes four to eight. Trade reports in the spring estimated that around half of global projects were facing delays because of power limits and equipment shortages and that more than ten gigawatts anticipated for 2026 had not yet begun construction.

The response of the large buyers has been to lock up power directly. In January, Meta agreed a twenty year power purchase arrangement with Vistra for more than 2,600 megawatts from three nuclear plants, one of several large nuclear and gas deals struck by the technology companies, and the International Energy Agency has been quoted as expecting global data centre electricity consumption to exceed a thousand terawatt hours by the end of 2026, roughly the annual use of Japan. I would treat the more precise power forecasts with caution, since they vary a great deal between sources and depend on assumptions about how efficient the chips become, but the direction is consistent: electricity has moved from a line item to a strategic constraint. This matters for investors because it changes who holds scarcity. In the first phase of the cycle, the scarce thing was the chip. In the second, it appears to be the ability to energise a building, and the companies that own power rights, interconnection positions and generation capacity have acquired a kind of leverage that has little to do with technology.

Now turn to the layer that I think is least understood outside the financial press: how all of this is paid for. For the first years of the cycle, the largest technology companies funded their spending from operating cash flow. That is changing. A report from Bloomberg's analysts, republished by an energy news service in February, counted at least 200 billion dollars of debt raised by AI related companies and projects in 2025, probably an undercount because many deals are private, and noted that Morgan Stanley expected 250 to 300 billion dollars of issuance from the large technology companies alone in 2026. Several junk bond deals totalling about seven billion dollars were sold last year to finance particular new data centres, and one data centre operator borrowed at roughly nine percent to refinance existing debt. JPMorgan projected that securitisations backed by data centres could reach 30 to 40 billion dollars a year in 2026 and 2027, and it later estimated that about 4.1 trillion dollars of AI related debt could be issued through 2030. The Federal Reserve Bank of Dallas has written that the financing needs are large and persistent enough to affect the supply of long duration bonds and therefore interest rates.

Two features of this financing deserve attention. The first is the growing use of structures that keep debt off the balance sheets of the technology companies. Instead of owning a data centre, a company may lease it from a developer that owns it through a special purpose vehicle, which borrows against the lease. That is a sensible way to spread risk, but it makes total exposure harder to measure, and credit analysts have been warning that lease commitments and project financings need to be added to reported debt to see the real picture. The second is the gap that Morgan Stanley's research has described between the roughly 2.9 trillion dollars of global data centre investment it expects through 2028, excluding power, and the roughly 1.4 trillion dollars that large technology companies can fund from their own cash flows. The difference of about 1.5 trillion dollars is a funding gap, not a forecast of debt, but it indicates the scale of outside money that the plans require, and a large share of it is expected to come from private credit and structured finance.

The connections between companies in this cycle deserve their own paragraph. The chip supplier has announced plans to invest up to 100 billion dollars in a leading model developer, which will in turn buy many gigawatts of the supplier's processors. A rival chip designer agreed to supply that developer with six gigawatts of computing in exchange for warrants over about ten percent of its shares. In March, the developer announced a funding round of about 122 billion dollars at a valuation of roughly 852 billion dollars, with reported participation from a large cloud provider at about 50 billion, the chip supplier at about 30 billion and a Japanese investor at about 30 billion, and trade reports say the developer has cumulative computing commitments of around 600 billion dollars through 2030, much of it on take or pay terms that require payment whether or not the capacity is used. The developer's chief executive has spoken of commitments totalling more than a trillion dollars, a figure that includes arrangements of different kinds and that analysts have urged readers not to simply add together.

I describe these arrangements without judgement, because they are legal and in many cases commercially rational. Suppliers who invest in customers are buying demand, and customers who grant warrants to suppliers are buying priority. But they have a consequence that anyone assessing the cycle should keep in mind: reported demand and reported revenue are partly funded by the suppliers themselves, which makes it harder to know how much comes from independent customers willing to pay. If demand is strong, this circularity is a feature of a young market. If demand weakens, the same links can transmit a problem from one company to several others in a hurry. The recent history of the flagship project associated with the developer is an illustration of the kind of adjustment that large plans undergo. The original joint venture announced in January 2025 with Oracle, SoftBank and the Abu Dhabi investor MGX was described as a 500 billion dollar programme to deliver ten gigawatts. By spring, trade reports said that plans to expand the flagship Texas site beyond its initial build had been dropped, the company announced on April 29 that the name would henceforth cover its entire long term compute effort with all partners and not the original venture, and analysts estimated that only a small fraction of the contracted capacity was actually plugged in. The company responded publicly to press reports of missed targets. None of this shows that the plans will fail, but it shows how far contracted capacity can run ahead of working capacity.

The next layer up is the laboratories that train the models. Funding data for the first quarter of 2026, published by Crunchbase on April 1, showed AI companies capturing about 242 billion dollars of roughly 300 billion dollars of global venture funding, around 80 percent, with a handful of companies absorbing most of it, and that quarter alone deployed about as much venture capital as had been invested across all of 2025 in other analyses. The costs these laboratories bear explain the scale of the rounds. Training a leading model and, above all, serving it to hundreds of millions of users requires computing on the order of billions of dollars a year, and reports suggest that at least one leading developer does not expect to generate positive cash flow before 2028. The business model is a race: to keep spending on computing at a rate that stays ahead of competitors while revenue grows fast enough to justify the next round. Prices to customers keep falling while the cost of the capacity needed to serve them keeps rising, a squeeze that favours the companies with the deepest financial backing and puts the most strain on the second tier.

Above the laboratories are the applications, which I covered in my study of the companies quietly changing business. The money going into them is smaller but arguably more informative, because it reflects whether customers are paying. Vertical companies in law, healthcare and customer service have reported revenue growing at rates that few software businesses have ever matched, though the precise figures vary by source. At the same time, the early months of 2026 brought a sharp sell off in the shares of established software companies, which investors began to treat as exposed to automation of their customers' work, a reaction that press accounts described as erasing very large amounts of market value. These two facts fit together: investors are paying a premium for companies that sit inside new workflows and a discount for companies whose value rested on the cost of building software or the number of human seats.

Putting the stack together, who is actually earning the returns? At the chip layer, the returns are enormous and audited. At the foundry, memory and packaging layer, they are strong and rest on scarcity. At the cloud providers, revenue is growing quickly but free cash flow is under pressure from spending that is rising faster than income, and borrowing is increasing. At the power and real estate layer, returns accrue to whoever owns interconnection rights and generation, and delays are costing developers. At the financing layer, lenders are earning interest and fees, with the risk concentrated in the less creditworthy borrowers. At the laboratory layer, revenue is growing fast but is dwarfed by spending, and the outcome depends on continued access to capital. At the application layer, the fastest growing companies are showing real revenue, though valuations are demanding. And the customers, in every case, are benefiting from falling prices and rising capability, which is the part of the story that rarely appears in an investment table. In the longer run, the customers' share of the surplus is the main reason the investment is justified.

Is this a bubble or a supercycle? I think the question is badly posed, because both words describe the same phenomenon at different time scales. History suggests that large infrastructure cycles are almost always right about the direction and wrong about the timing and the price. The telecommunications boom of the late nineties laid fibre that was used productively for decades, and it also bankrupted many of the companies that financed it. There are reasons to think this cycle differs in both directions. The equipment depreciates faster than fibre did, since each new chip generation makes the last one less valuable, which raises the risk for those who borrow against it. But demand is visible today in a way it was not for fibre, with the largest buyers saying they are constrained by supply, and the supplier's own order book supporting that. My view is that some parts of the stack will earn handsome returns for years, some will be overbuilt and repriced painfully, and that the dividing line will be between businesses funded by cash flow and businesses funded by borrowing against assumptions. I cannot say which will be which, but I would look hardest at the financing structures.

The macroeconomic setting adds a note of caution. The International Monetary Fund's April projection put global growth at 3.1 percent for 2026, below the pace of recent years, and the war in the Middle East that began at the end of February has raised energy and shipping costs and unsettled inflation. A technology investment boom running against a slower and more uncertain global economy is not unusual, but it raises the stakes on the assumption that AI revenue will rise quickly enough to carry the financing. If borrowing costs rise or credit appetite falls, the most leveraged parts of the stack would feel it first, which is another reason to watch the financing layer closely.

Now for my own predictions, stated as judgement and not as forecasts from any institution, and written from the vantage point of the first of July without knowledge of what the next quarter will bring. Over the next six to twelve months I expect spending plans by the largest buyers to be revised up again, unless the economy or financing conditions deteriorate sharply. I expect debt issuance related to AI to keep rising and for investors to differentiate more sharply between strong and weak borrowers, so that the cost of capital for leveraged builders diverges from that of the largest companies. I expect power and equipment constraints to delay a meaningful share of announced capacity, which will keep prices for available capacity firm. I expect memory and packaging to remain tight, and for memory to become a larger share of the cost of an AI system. I expect continued consolidation among second tier model developers and specialist cloud operators, through acquisition or quiet failure. I expect prices for model usage to keep falling, which will favour application companies and squeeze anyone selling raw capability. And I think it is more likely than not that at least one leveraged participant in the build out will be forced into a restructuring within a year, without that event changing the direction of the larger cycle.

What does this mean for a founder in Bangladesh? For most of us the build out is something that happens elsewhere and affects us through prices and access. The first implication is encouraging: the intelligence that powers a product is getting cheaper, which lowers the cost of building useful software for a local market. The second is a caution: that intelligence is priced in dollars and supplied by a small number of foreign companies, so a product that depends on it inherits exchange rate risk and supplier risk, and should be built to switch providers. The third is that the physical constraints of the build out, the power and the packaging, suggest that capacity will remain tight and that prices for the most advanced services may not fall as fast as those for simpler ones. The fourth is that local power conditions make hosting heavy computing at home unrealistic for now, so the practical strategy is to build applications on top of global infrastructure, and to compete on knowledge of local language, law, payments and customers, which no foreign laboratory will supply. The money has mostly gone into the lower layers of the stack. The opportunities for small companies are in the top layer, and in the gaps between the layers where someone has to make everything work for a particular customer.

It is worth listing the indicators I would watch to understand where the cycle is heading. I would compare each large buyer's capital spending with its cloud revenue and watch whether the gap narrows. I would track the supplier's order outlook and its mix of customers, especially the share coming from outside the largest cloud providers. I would watch the prices for renting computing capacity, since falling prices signal oversupply and rising prices signal scarcity. I would follow credit spreads on bonds issued to finance data centres, and the terms on private credit deals, because lenders tend to see problems before equity investors do. I would watch the share of announced data centre capacity that is actually energised. I would watch the prices of high bandwidth memory and advanced packaging. And I would watch whether independent customers, not only suppliers' affiliates, are signing large commitments.

The limits of this study are plain. Most of the largest figures are plans or estimates that analysts have produced from partial information, and they differ between sources. The audited numbers are few, mostly from Nvidia and the other public companies, and the private companies' finances are known only through reports. Several of the claims about supply chain shares and capacity come from industry analysts and not from the companies. Anything I have said about the future is a judgement. The period I describe ends in the last days of June 2026, and events since then, of which I have no knowledge, may have changed the picture. Anyone relying on a number here should go back to its source, check its definition and its date.

The short version is this. The money has gone into a stack. At the bottom, chips earn extraordinary margins because they are scarce. Around the chips, the foundry, memory and packaging suppliers earn strong returns for the same reason. Above them, buildings and power have become the real bottleneck, and the cost of connecting to the grid has become a strategic asset. Financing has shifted from cash flow towards debt, leases and private credit, and much of the demand is linked in circles of suppliers investing in customers. The laboratories burn cash to stay in the race, and the application companies show the revenue that makes the whole thing plausible. The customers, meanwhile, capture much of the benefit through falling prices. Whether the cycle ends as a durable platform or a painful overbuild will depend less on the technology than on the financing, which is why the question I would ask of any company in this stack is not how impressive its product is but who is paying for its next year, and what happens if they stop.

Sources: NVIDIA, financial results for the first quarter of fiscal 2027 (https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000051/q1fy27pr.htm) and fourth quarter and fiscal 2026 (https://www.sec.gov/Archives/edgar/data/1045810/000104581026000019/q4fy26pr.htm); NVIDIA first quarter 10-Q (https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000052/nvda-20260426.htm); Futurum Group, AI capex 2026 (https://futurumgroup.com/insights/ai-capex-2026-the-690b-infrastructure-sprint/); Goldman Sachs, Why AI Companies May Invest More than $500 Billion in 2026 (https://www.goldmansachs.com/insights/articles/why-ai-companies-may-invest-more-than-500-billion-in-2026) and Tracking Trillions (https://www.goldmansachs.com/insights/articles/tracking-trillions-the-assumptions-shaping-scale-of-the-ai-build-out); EnergyNow on AI data centre debt (https://energynow.com/2026/02/the-3-trillion-ai-data-center-build-out-becomes-all-consuming-for-debt-markets/); Federal Reserve Bank of Dallas, AI debt financing and duration supply (https://www.dallasfed.org/research/economics/2026/0210-searls-aifinancing); Bloom Energy, 2026 Data Center Power Report (https://www.bloomenergy.com/wp-content/uploads/2026-power-report.pdf); ATK Energy on grid interconnection (https://atkenergygroup.com/blog/grid-interconnection-data-centers/); Next Financial on CoWoS capacity (https://nextfinancial.substack.com/p/the-60-problem); Silicon Analysts on foundry allocation (https://siliconanalysts.com/analysis/foundry-allocation-status-q1-2026); Data Center Dynamics on Stargate (https://www.datacenterdynamics.com/en/analysis/openai-building-stargate-nvidia-oracle-chatgpt/); OpenAI on Stargate (https://openai.com/index/building-the-compute-infrastructure-for-the-intelligence-age/); Digital Applied on first quarter 2026 AI venture funding (https://www.digitalapplied.com/blog/ai-venture-funding-2026-where-242b-went-data-atlas); IMF, World Economic Outlook, April 2026 (https://www.imf.org/en/publications/weo).