- Strategy
- Execution
- People
- Finance
- Startups
- Digital Transformation
- CXO Advisory
- Latest Trends
- Lifestyle
- Events
Everyone now agrees AI is a bubble. That is the least useful thing you can know about it. The useful questions are: what cracks first, who absorbs the loss, what becomes cheap, and where the real returns show up afterwards. History has run this experiment three times. The answer is uncomfortable for Silicon Valley and very good for the CFO who is paying attention.
On 18 September 1873, the Philadelphia bank of Jay Cooke & Company stopped honouring withdrawals. Cooke had been the most trusted financier in America, the man who sold the Union's war bonds. He had also tied the firm to the Northern Pacific Railroad, a transcontinental line being built across land almost nobody lived on yet. The New York Stock Exchange closed for ten days. The United States entered the longest depression in its history.
The railways did not disappear. The track stayed in the ground, got bought out of receivership for cents on the dollar, and over the next two decades carried the wheat, steel and mail-order catalogues that built the modern American economy. The investors were wiped out. The infrastructure won.
That is the pattern I want to walk you through, because it is the pattern we are in. The C-suite question is not "is AI a bubble". It is "which side of the 1873 ledger will my company be on".
Four companies are spending roughly the GDP of Sweden on data centres this year. Amazon, Alphabet, Microsoft and Meta have guided to $720–745 billion of 2026 capital expenditure between them, and Moody's expects the figure to approach $1 trillion in 2027. In the first quarter of 2026, AI-related data centre and software investment accounted for about three-quarters of all US GDP growth. Consumer spending contributed less than a fifth. The American economy is, for the moment, an AI construction site.
The money is increasingly circular. Nvidia holds a $30 billion equity stake in OpenAI. OpenAI has committed to about $300 billion of computing capacity from Oracle over five years. Oracle buys its chips from Nvidia. AMD has agreed to supply OpenAI with 6 gigawatts of GPUs and handed it warrants for up to 160 million AMD shares. The Bank for International Settlements has called the web of commitments "opaque and interconnected", and estimates of the interlocking total run to $1–1.5 trillion. When your biggest customer is also your shareholder's shareholder, revenue stops being an independent signal.
The revenue to justify it does not yet exist. Bain estimates that the industry needs $2 trillion of annual AI revenue by 2030 to fund the compute being built, and is $800 billion short even after counting every projected AI saving. OpenAI, the sector's bellwether, ended September 2026 at a $50 billion run rate, some $18–20 billion below the figure investors had been circulating, with projected cash burn of $278 billion through 2030.
The cost of money has turned. The Federal Reserve raised rates in September 2026 for the first time since 2023, the US 10-year yield sits above 5.2%, and futures price another hike in December. Oracle carries about $132 billion of net debt against negative free cash flow. CoreWeave carries $46 billion at 8.2 times EBITDA, paying up to 9.75% on its notes. Bubbles do not burst because people stop believing. They burst because the carry gets too expensive.
And the market has noticed. On 8 October the Nasdaq had its worst session since August: CoreWeave fell 7.8%, Oracle 5.5%, Nvidia 2.9%. Chip stocks had gained more than 80% in 2026 before that day. One bad print on OpenAI's revenue was enough to take 5% off the world's second-largest software company. That is not a market that believes its own story. It is a market looking for the exit.
Most commentary treats the AI bubble as a valuation problem. It is actually three separate problems that happen to be sitting on top of each other, and the crack will run along whichever gives way first.
The depreciation fault line. A GPU is not a railway sleeper. Track laid in 1873 was still carrying freight in 1920. A data centre GPU bought in 2024 is economically obsolete in three to four years, and Nvidia ships a better one every twelve months. Yet several hyperscalers have quietly extended the accounting useful life of their servers to five or six years. That lowers reported depreciation today and flatters margins. It also means the balance sheets are carrying hundreds of billions of dollars of hardware at values the resale market will not honour. When the write-downs come, they will come all at once, because every CFO is using the same schedule.
The customer fault line. Who ultimately pays for $745 billion of capex? Not consumers paying $20 a month. It has to be enterprises, and enterprises are not yet paying. MIT's NANDA initiative found that 95% of corporate generative AI pilots deliver no measurable P&L impact. I have sat in enough boardrooms this year to confirm it: the pilots are everywhere, the budget line is still "innovation", and the CFO has not yet been asked to fund it from operating margin. The day she is asked, the demand curve that justifies $1 trillion of 2027 capex gets a lot flatter.
The financing fault line. The 1873 crash was not caused by too few passengers. It was caused by Jay Cooke being unable to sell the next tranche of Northern Pacific bonds. The dotcom crash was not caused by too few web users. It was caused by Global Crossing and WorldCom being unable to roll their debt. Every infrastructure bubble ends at the financing layer, and this one has a financing layer unlike any before it: a neo-cloud sector borrowing at near-junk rates to buy chips from the company that is also its investor, to serve a customer that burns more cash than any private company in history. One failed debt raise, one covenant breach, one downgrade, and the circle stops turning.
The railway bust needed a single bank to fail. The dotcom bust needed a single accounting scandal. The AI bust needs one of three things to go wrong, and all three are already under strain.
The losses arrive in a sequence, and the sequence matters because it tells you how long you have.
The first casualties are the leveraged intermediaries: the neo-clouds and the data centre developers who borrowed to build capacity for a single anchor tenant. In 1873 these were the regional railroads with one bond issue and one route. In 2001 they were Global Crossing, a $47 billion company that filed for bankruptcy with $25.5 billion of assets, and the sixty-odd telecoms that followed it. Today they are the companies paying 9% coupons on GPU-backed debt. They go first, and they go fast, because a tenant renegotiating one contract is enough.
The second casualties are the equity holders in the model labs. OpenAI is seeking $30 billion at a $1.4 trillion valuation while burning $12 billion a quarter. If the next round prices flat or down, every late-stage investor and every employee holding paper from 2025 is underwater. This is where the "who loses their savings" question lands: not in public markets first, but in the pension funds, sovereign wealth funds and family offices that bought into the last three private rounds. Gulf capital is heavily represented here. More on that below.
The third casualties are public shareholders broadly. Microsoft, Alphabet, Amazon and Meta are about 17% of the S&P 500. Add Nvidia and Broadcom and it is 28%. A 30% correction in six stocks is an 8–9% hit to every index fund, every 401(k), every NRI's US brokerage account. This is the phase that gets the headlines, but it is the least dangerous for operating companies, because by then the useful part has already started.
The last casualties are the hyperscalers themselves, and they mostly survive. Amazon, Microsoft and Alphabet have advertising, retail and software cash flows that can absorb a $100 billion write-down. They did in 2001 and 2008. They will emerge owning the infrastructure, exactly as J.P. Morgan emerged from 1873 owning the railroads.
After every infrastructure crash there is a window, usually two to four years, when the assets built at the peak trade for a fraction of their cost. This is the single most profitable moment in the cycle, and almost nobody who lived through the boom is positioned to take it.
By 2002, roughly 97.5% of the long-haul fibre laid during the dotcom boom sat dark, carrying nothing. Creditors sold the bankrupt networks for less than they had cost to trench. Level 3 Communications bought up that capacity and at its peak carried about 70% of global internet traffic. CenturyLink paid $34 billion for it in 2017. The fibre nobody wanted in 2002 is the fibre your Zoom call runs on today.
The AI equivalent will be distressed compute. Expect three things in the clean-up window. First, GPU rental prices fall 60–80% as neo-cloud creditors liquidate inventory, and the "compute is scarce" premium that has defined 2024–2026 evaporates. Second, half-built data centre shells in Texas, Virginia, Jamnagar and Abu Dhabi get sold with their grid connections and cooling in place, which are the genuinely scarce assets, to whoever has cash. Third, the model labs that survive do so by consolidating: expect two or three foundation model companies where there are now eight, with the hyperscalers absorbing the rest the way Morgan absorbed the railroads.
For the C-suite, the clean-up is the buying window. The companies that locked in multi-year AI contracts at 2025 prices will be paying triple the market rate. The companies that waited, built internal capability and kept dry powder will be able to buy enterprise AI capacity at commodity prices. The lesson of Level 3 is that the fortune goes to the balance sheet that is liquid when everyone else is forced to sell.
Here is the part the bubble commentary always misses. The productivity gains from a general-purpose technology do not arrive during the boom. They arrive five to fifteen years after the crash, once the technology is cheap, boring and embedded in processes nobody writes about.
The railways' real economic payoff was not the transcontinental lines that bankrupted their builders. It was the branch lines, the standardised freight rates and the Sears catalogue, which only became possible once shipping a parcel from Chicago to a farm in Kansas cost next to nothing. The internet's payoff was not Pets.com. It was the unglamorous decade from 2003 to 2013 when cheap bandwidth let a company in Bangalore run a back office for a bank in Charlotte, and the Indian IT services industry grew from $10 billion to over $100 billion.
AI's equivalent is already visible if you look away from the chatbots. It is a mid-sized logistics company in Pune using a $200-a-month model to reconcile invoices that used to take four clerks. It is a property developer in Dubai running due diligence on a hundred title deeds in an afternoon. It is a hospital group in Hyderabad coding insurance claims with 40% fewer rejections. None of these need a frontier model. All of them need compute at commodity prices, which is precisely what the crash delivers.
Gartner's hype cycle names this stage the plateau of productivity, and it is where the money is made by users rather than vendors. The roll-out phase rewards three kinds of company: the ones with proprietary data that cheap models can finally be pointed at; the ones with the process discipline to re-engineer a workflow rather than bolt a chatbot onto it; and the ones that kept their AI spend variable rather than committed. If your company has those three things, the bubble bursting is the best thing that could happen to your margins.
India has committed to building the infrastructure at the exact moment the US is debating whether it was overbuilt. Reliance announced a $110 billion, seven-year AI infrastructure plan in February, with gigawatt-scale campuses under construction in Jamnagar. Adani followed with $100 billion. Google has committed $15 billion to a Visakhapatnam hub. The government expects more than $200 billion of AI infrastructure spending over two years. Tata is building OpenAI's Indian capacity.
This is either very good timing or very bad, and the difference comes down to one variable: cost of power. Reliance is building on 10 gigawatts of its own surplus solar. Adani owns its generation. A data centre whose electricity costs a third of Virginia's can survive a 70% collapse in GPU rental prices; a leveraged US neo-cloud buying grid power at market rates cannot. If the Indian builders finance conservatively and stay on their own power, they are positioned to be the Level 3 of this cycle: the low-cost owner of capacity when the roll-out phase arrives. If they finance it the way Jay Cooke financed the Northern Pacific, they will be the first casualty on this side of the world.
The second Indian exposure is the IT services industry, and it is already in the crash phase. TCS's revenue fell 0.5% in FY26, its first annual decline; HCL's CEO has warned of 3–5% "AI deflation" on revenue; the sector's $250 billion of annual exports is built on billing hours that AI compresses. The irony is sharp. Indian IT was the greatest beneficiary of the last infrastructure bust, when cheap fibre made offshoring possible. It is now the most exposed incumbent to this one. The CEOs of the big four have perhaps three years to turn deflation on their core business into a share of the roll-out economy, and the clock started in April.
For the Gulf, the exposure is on the capital side. Sovereign funds and family offices from Abu Dhabi, Riyadh and Doha are among the largest backers of the late-stage model lab rounds and the data centre developers. These are the equity holders who absorb phase one. The DIFC conversations I am having now are not about whether to invest in AI; they are about whether the 2025 vintage was the peak. For NRI investors in the Gulf with US index exposure, the practical question is simpler: six stocks are 28% of your S&P 500 fund, and you did not choose that concentration.
Intellectual honesty requires naming the exit from the analogy, and there are three.
First, this bubble is funded differently. The railways and the fibre networks were built on debt sold to the public. Most AI capex is funded from the operating cash flow of four companies that each generate more than $100 billion a year. A bubble financed by retained earnings deflates; it does not detonate. The neo-cloud debt is the exception, and it is large, but it is not the whole system.
Second, enterprise revenue could arrive faster than Bain's model assumes. OpenAI's enterprise run rate reportedly grew 107% in the third quarter of 2026. If that holds for two more years, the $800 billion gap closes on its own. I think the 95% pilot failure rate argues against it, but I have been wrong about adoption speed before.
Third, the chips may hold their value longer than the pessimists expect. If model efficiency gains slow and inference demand keeps compounding, a 2024 GPU may still be earning its keep in 2029, and the depreciation fault line never opens. Watch two things to judge this: the second-hand price of H100s, and whether any hyperscaler shortens its server useful-life assumption. The first is the market's verdict; the second is the auditors'.
If all three hold, this is 1920s electrification rather than 1873: a boom that deflates slowly into ubiquity without a crash. That is possible. It is not what the financing structure, the revenue gap or the bond market are currently telling us.
None of this is a reason to stop using AI. It is a reason to stop buying it the way the vendors want you to. Here is what I would put on the agenda before the year-end board meeting.
Make every AI commitment variable. No multi-year reserved-capacity contracts, no prepaid token bundles, no three-year enterprise licences priced at 2025 scarcity. Compute will be 60–80% cheaper in the clean-up window. Buy monthly.
Audit the pilots against P&L, not enthusiasm. Ask one question of every AI initiative: which cost line or revenue line moves, by how much, by when? Kill the ones that cannot answer. MIT's 5% that work share a trait: they were owned by a business unit with a number to hit, not by an innovation team.
Build the data asset now. The roll-out phase rewards proprietary data pointed at cheap models. Cleaning, structuring and governing your own data is the one AI investment that appreciates through a crash.
Check your concentration. Your treasury portfolio, your pension fund and your personal holdings are likely 25–30% exposed to six AI stocks through index funds alone. Decide deliberately whether that is the exposure you want.
Prepare the buy list. Which capabilities, teams or distressed assets would you acquire if prices fell 70%? The CEOs who made fortunes in 2002 and 2009 had the list written before the crash. Write yours in Q4 2026, not Q2 2027.
The bubble will burst. The infrastructure will stay. The question for every executive reading this is the same one Jay Cooke's depositors faced in September 1873: are you the one holding the bonds, or the one buying the railway?
Is AI really a bubble, or just an expensive boom? The honest answer is that the technology is real and the financing is a bubble. Four hyperscalers are spending $720–745 billion on AI capex in 2026, while Bain estimates the industry is $800 billion short of the annual revenue needed to pay for it by 2030. Booms are funded by profits; bubbles are funded by the expectation of profits. The AI build-out is currently the second kind.
When will the AI bubble burst? Nobody can date it, and this article does not try. What history shows is the trigger: infrastructure bubbles end at the financing layer, when a leveraged builder cannot roll its debt or a flagship cannot raise its next round. Watch OpenAI's next funding round, CoreWeave's refinancing, and any hyperscaler shortening its server depreciation schedule. Those are the three tripwires.
Who loses money when it bursts? In sequence: lenders to leveraged data centre developers and neo-clouds; late-stage private investors in model labs, including Gulf sovereign funds, pension funds and employees holding 2025-vintage paper; then ordinary index investors, since six AI stocks make up roughly 28% of the S&P 500. The hyperscalers themselves mostly survive and end up owning the infrastructure.
What becomes cheap afterwards? Compute. In the dotcom clean-up, 97.5% of long-haul fibre sat dark and was sold for less than it cost to lay. The AI equivalent is GPU capacity: expect rental prices to fall 60–80% as creditors liquidate inventory, and half-built data centre shells to change hands for the value of their grid connection. Enterprise AI that costs $1 million a year today will cost a fraction of that in the clean-up window.
Will AI itself fail or disappear? No. The railways survived 1873, the fibre survived 2001, and the models and chips will survive this. What disappears is the premium: the scarcity pricing, the trillion-dollar valuations and the assumption that whoever spends most wins. The technology moves from the vendors' balance sheets to the users' P&Ls, which is where the productivity gains have always shown up.
How does this compare with the dotcom crash? The shape is similar, the funding is different. Dotcom infrastructure was built on public debt and IPO money, so the crash detonated. Most AI capex is funded from the operating cash flow of four companies earning over $100 billion a year each, so the core may deflate rather than explode. The neo-cloud and model lab layer, however, is financed almost exactly like the 1990s telecoms, and that is where the detonation risk sits.
What does the AI bubble mean for India? Two opposite exposures. India is committing more than $200 billion to AI infrastructure, led by Reliance's $110 billion and Adani's $100 billion plans; if financed conservatively and run on their own cheap power, these builders could become the low-cost owners of capacity when the roll-out phase arrives. Meanwhile Indian IT services, the biggest winner of the last infrastructure bust, is the most exposed incumbent this time: TCS's revenue fell in FY26 and HCL is guiding for 3–5% "AI deflation".
What does it mean for Gulf investors and NRIs? Gulf sovereign and family-office capital is concentrated in the late-stage private rounds that absorb phase-one losses. NRIs with US index funds carry a 25–30% exposure to six AI stocks they did not consciously choose. In both cases the question is not whether to own AI, but whether the 2025–2026 vintage was the top.
Should my company stop investing in AI? No. It should stop buying AI the way vendors want it bought. Keep spend variable, kill pilots that cannot name a P&L line, invest in your own data, and keep cash for the clean-up window. The companies that did this in 2002 and 2009 bought their competitors' infrastructure for cents on the dollar.
What is the single most important thing a CXO should do this quarter? Write the buy list. Decide now which capabilities, teams and distressed assets you would acquire if prices fell 70%, and what balance sheet you need to be able to act. The executives who made fortunes after 1873, 2001 and 2008 had the list written before the crash, not after.
Ideas:
Is AI in a bubble in 2026? Four hyperscalers are spending $720–745bn on AI capex this year against an $800bn gap between the revenue that compute needs and the revenue AI generates. Bain's analysis says the industry needs $2trn of annual AI revenue by 2030. Most analysts now treat a correction as a question of when, not if.
What happens when the AI bubble bursts? History suggests three phases: a crash that hits leveraged data centre developers and late-stage private investors first, a two-to-four-year clean-up where compute prices fall 60–80% and assets consolidate, then a roll-out phase where cheap AI drives productivity in ordinary businesses.
Who loses money if the AI bubble bursts? In order: neo-cloud debt holders, late-stage equity investors in model labs (including Gulf sovereign funds and pension funds), then broad index investors, since six AI stocks are about 28% of the S&P 500.
What should a CEO or CFO do about the AI bubble? Keep AI spend variable rather than committed, audit pilots against P&L, invest in proprietary data, check portfolio concentration, and prepare an acquisition list for the clean-up window.
