Red Teaming The AI Investment Thesis
Testing Strength Through Strategic Attack; Uncovering Weaknesses
This post was written in collaboration with Kevin Koharki (MBA, PhD), Associate Professor of Finance and Accounting at Purdue University and founder of CAE Consulting. Koharki advises Fortune 100 companies across industries including banking, insurance, aerospace, and defense. He is a frequent keynote speaker and financial analyst whose insights have been featured in publications including the Wall Street Journal and the Harvard Business Review. (https://www.kevinkoharki.com/)
Testing The AI Investment Thesis
AI will change the world. Of that there can be little doubt. That’s precisely what has investors so excited.
The danger is that periods of maximum optimism often coincide with minimum scepticism. As enthusiasm builds, investors become less inclined to challenge the assumptions underpinning the investment case.
Let’s do the opposite.
Let’s red team the AI investment thesis.
By deliberately playing devil’s advocate, we’ll pressure test the assumptions, challenge the consensus and search for weaknesses before they become obvious to everyone else.
We’ll ask the questions investors should be asking, but perhaps aren’t:
Can every participant in the AI value chain really be winning at the same time?
Are today’s reported profits an accurate reflection of the underlying economics?
Can the unprecedented capital expenditure being committed realistically earn an acceptable return?
How much of today’s AI demand represents durable economic value, and how much is simply experimentation?
Who ultimately bears the cost of the AI investment boom?
Will AI’s economic benefits arrive as quickly as investors currently expect?
When the capital cycle turns, who will still be standing?
Let’s begin.
Is It Possible To Have Winners Without Losers?
There’s something unusual happening in the AI economy.
Everywhere you look, somebody appears to be making money.
Upstream, AI hardware suppliers are reporting revenues that look like the GDP of a small country and they maintain margins so thick you could cut them with a knife.
Downstream, the hyperscalers are engaged in an AI infrastructure arms race, spending hundreds of billions of dollars as though money is going out of fashion. The market has not seen CAPEX spending like this since the dot com boom at the turn of the millennium.
Yet, despite this unprecedented capital expenditure, reported earnings show remarkably little sign of deterioration.
Debt markets continue to provide abundant financing, while equity investors seem happy allocating capital to all participants in the value chain.
Everyone appears to be winning.
That alone should make investors uncomfortable.
Competitive markets rarely produce only winners. One company’s exceptional profits are usually another company’s exceptional costs.
For instance, during the pandemic, shipping companies prospered while importers paid the bill. Then came the post-Covid semiconductor shortage when chipmakers benefited from higher prices while automakers absorbed the pain.
Gains don’t materialize out of nothing and economic costs don’t disappear. Yet today’s AI economy appears different. At first glance, there doesn’t appear to be a loser anywhere in the system.
That raises an obvious question.
How can everyone be winning at the same time?
Artificial intelligence is transforming the economy. That much seems increasingly difficult to dispute. The arrival of ChatGPT marked an inflection point, making it clear that AI would reshape the way we work, create and allocate capital.
Our destination became obvious.
That didn’t make the journey any less perilous.
Throwing money at an innovative new technology is not necessarily a recipe for success. It’s often possible to be right about the ‘what’, but wrong about the ‘when’.
History is full of examples where investors correctly identified transformational technologies yet generated disappointing returns. Railways, radio, airlines, telecommunications and the internet all changed the world. But they experienced periods where capital flowed into the technology much faster than economic returns flowed back out.
The greatest investment risks rarely arise because a technology fails. They arise because reality, however impressive, fails to match what investors have already priced in.
The more I examined today’s AI economy, the more one conclusion stood out.
Several timing mismatches are unfolding simultaneously.
Accounting recognises revenues before many of the associated costs. Capital is being deployed years before its returns can be measured. Infrastructure is being built faster than organisations can fully utilise it. Productivity gains will likely emerge gradually, while many of the costs are already flowing through the economy.
None of this implies AI will fail.
It may prove to be one of the most important technologies ever developed.
But timing matters. For investors, understanding these timing mismatches may ultimately prove just as important as understanding the technology itself.
The Accounting Anomaly
One of the more interesting features of the AI investment boom has nothing to do with artificial intelligence.
Instead, it’s a quirk of modern accounting.
Current accounting standards recognise the benefits and costs of the AI build-out in different reporting periods. The result is a temporary distortion where reported profitability can look considerably stronger than the underlying economics.
The mechanics are surprisingly simple.
When a hyperscaler purchases $30 billion of GPUs, the supplier’s income statement, balance sheet, and statement of cash flows light up like a Christmas tree.
But on the other side of the transaction, the buyer reports something very different.
The investment is recorded on the balance sheet as ‘Construction in Progress’ (CIP), and on the cash flow statement as capital expenditure, yet there is no sign of it on the income statement.
The accountants effectively say, “You haven’t turned this on yet; you haven’t deployed it to serve customers; so we won’t force you to start bleeding depreciation just yet.”
That distinction matters.
The supplier recognises the economic benefit today, while the buyer recognises much of the economic cost sometime later.
For a period of six, twelve or even eighteen months, those chips sit in warehouses or are slowly racked into data centers, silently awaiting activation. So while the vendor’s top line soars, the buyer’s earnings remain pristine.
Investors naturally anchor on today’s earnings. Less attention is paid to the costs that have already been committed but have yet to appear in the income statement.
This observation sits behind comments from investor Michael Burry, who has argued that reported earnings at companies such as Meta and Oracle materially overstate their underlying economics by tens of billions of dollars.
Whether Burry’s precise calculations prove correct is almost beside the point. The broader principle is what matters.
Large amounts of capital have already been committed, while much of the associated expense still lies in the future. Burry spotted it, while most AI investors have yet to connect the dots.
We are currently living in a high-stakes experiment in deferred economic gravity. The books will balance in the end, they always do, but the timing of that balancing act is everything. When it’s finally achieved, it will separate the companies that built genuine economic value from those that simply moved silicon around a balance sheet.
It is important to stress that none of this is improper. The accounting treatment is entirely in accordance with GAAP.
At an individual company level, it reflects the matching principle by recognising costs only once the assets begin generating revenue. But at a macro level across the economy as a whole, it creates a temporary illusion.
Although the costs are currently invisible, they haven’t disappeared.
Eventually, the income statement benefits of this timing mismatch begin to unwind.
The Deferred Depreciation Wave
As soon as CIP assets are placed into service, the depreciation axe starts to swing, cutting through income statement earnings with brutal regularity. It becomes a permanent feature of the cost base for the remaining life of the asset, and then for the assets that ultimately replace them.
As successive data centres are completed, depreciation arrives not as a single event but as a rolling wave of costs.
Hyperscaler margins currently sit well above their long-term averages, but when recent CAPEX spending becomes a depreciation expense, that is highly likely to change. Tomorrow’s earnings will increasingly reflect the cost of maintaining today’s investment.
Lets start by distinguishing General Purpose Technology (GPT) revolutions of the past from the current ArtificiaI Intelligence GPT. Railways, electricity grids, pipelines and fibre optic networks often remain economically productive for decades. GPUs don’t.
Technological progress is relentless. Every new generation of semiconductor typically delivers better performance, greater energy efficiency and lower cost per unit of computation. Hardware that appears state of the art today may become economically inferior only a few years from now.
The debate about whether the useful life of these assets is four years or six ultimately doesn’t matter.
The hardware will require replacement, of that there can be no doubt.
Accounting can influence when costs appear, but it can’t eliminate the economics.
Management therefore faces a difficult judgement.
Depreciate assets too aggressively and current earnings suffer. Depreciate them too slowly and the risk of future write-downs increases if technological obsolescence arrives sooner than expected.
Before concluding that a depreciation wave inevitably creates a financial crisis, several important qualifications are worth recognising.
Investors increasingly focus on EBITDA and cash generation rather than GAAP earnings alone. Accordingly, many analysts will simply shrug their shoulders and dismiss higher depreciation as little more than a non-cash accounting charge.
Under normal circumstances that view is entirely reasonable, but not in this case.
Depreciation and maintenance capital expenditure usually converge. On the cash flow statement, one is added back while the other is deducted, leaving free cash flow as a reasonable proxy for the true underlying economics of the business.
But the current AI investment cycle breaks that relationship.
Today, hyperscalers are investing extraordinary sums in new infrastructure. Cash leaves the business immediately, depressing free cash flow, while earnings and operating cash flow remains unscathred since much of the depreciation remains deferred as explained above.
The five largest hyperscalers now see their CAPEX exceeding operating cash flows.
As a result of deteriorating unit economics caused by capital intensive AI spending, the Maginificent-7 are suddenly underperforming other segments of the market.
The top five weighted companies in the S&P500 (Nvidia, Apple, Alphabet, Microsoft and Amazon) peaked at just over 30% of the index last year, but have since fallen back to 2023 levels of ~27%.
If this marks the peak of the mega-cap investment cycle and we are now pivoting towards mean reversion, there is certainly a long way to fall. The chart below measures the top 10% of stocks by size versus the entire US stock market and the long term average is well below current levels.
This can be broken down further to look at the top 10 largest stocks which currently account for 40% of the value of an index of 500 companies. That is significantly above where they were during the TMT (dot com) era.
Eventually, the build-out phase gives way to a digestion phase and construction slows (more on this shortly).
At this juncture, capital expenditure falls sharply, while depreciation rises as existing assets enter service.
The result is an unusual combination. Reported earnings weaken while free cash flow improves.
Cash conversion rates spike, and at first glance, investors may conclude these businesses have become highly cash generative.
The reality, however, is more nuanced.
Lower capital expenditure doesn’t necessarily imply lower economic costs. It may simply reflect a pause in the investment cycle.
Cash conversion improves because of accounting timing, not necessarily because the underlying economics have improved.
That makes valuing hyperscalers during the AI revolution unusually difficult.
Traditional valuation models assume a reasonably stable relationship between investment, depreciation and future cash generation.
The AI build-out temporarily disrupts all three.
Which brings us beyond accounting and into economics.
Creating value for society through artificial intelligence is one challenge. Capturing enough of that value to justify trillions of dollars of investment is another entirely.
As John Maynard Keynes observed in The General Theory of Employment, Interest and Money: “There is no clear evidence that the investment policy which is socially advantageous coincides with that which is most profitable.”
That insight sits at the heart of the bear case in relation to today’s AI investment boom, and that takes us directly to the AI capital cycle.
The AI Capital Cycle
The accounting timing mismatch explains why today’s earnings look unusually strong, but it’s the capital cycle that explains why they may not stay that way.
In a normal market, higher input prices eventually curb demand and restore equilibrium. Today, however, soaring semiconductor and memory chip prices appear to have had little effect on AI capital spending.
The incentives are easy to understand.
No management team wants to discover that demand has exceeded its available capacity while a competitor captures the opportunity instead. The cost of building too late may prove far greater than the cost of building too early.
Collectively, however, rational decisions can produce irrational outcomes.
In AI, the addressable markets of the hyperscalers overlap significantly. Every company is investing as though it can capture the majority of future demand, which is highly unlikely. In the past, these hyperscalers stuck to the swim-lane in which they dominated: Google in search, Microsoft in operating systems and enterprise software, Meta in social media, and so on. This time is different. They are all competing with each other in AI. Winner takes all, or winner takes most, is highly unlikely this time around.
Goldman Sachs estimates roughly $7.6 trillion of AI capital expenditure between 2026 and 2031, while Morgan Stanley estimates that the four largest hyperscalers will invest over $1 trillion during 2027 alone.
Much of this investment is not responding to existing demand. It’s a bet that future revenues will eventually justify today’s spending.
The economics are demanding.
A $1 trillion investment earning a 10% annual return requires $100 billion of profit. At a 12.5% operating margin, that implies roughly $800 billion of additional annual revenue. To put that in perspective, collectively the hyperscalers generated around $1.67 trillion of revenue last year. AI therefore needs to increase revenue by almost 50% within a remarkably short period, assuming today’s margins remain intact.
Those assumptions become even more demanding if AI ultimately proves to be a structurally lower margin business because of its capital intensity.
Even if AI adoption does accelerate dramatically, there may still not be enough revenue to justify the aggregate investment being made by these companies.
Meta has already acknowledged that portions of its compute infrastructure are sitting underutilised and it is actively seeking to sell its excess capacity.
We’ve seen this movie before. Railways, steel, shipping, telecommunications, solar panels and semiconductors all experienced investment booms that eventually created excess capacity. It would be naive to assume AI will escape the same capital cycle dynamics.
Eventually, every hyperscaler is likely to reach the same conclusion. The next data center becomes less urgent than improving the utilisation of the ones already built. When that happens, spending will slow and attention will shift from buying new semiconductors and memory chips to utilizing those already sitting on the balance sheet labelled CIP.
This is where the digestion phase begins. It’s a shift from expansion towards optimisation. Capital expenditure moderates and the economics change across the entire value chain.
The hardware vendors discover that their largest customers no longer need to expand capacity at the same pace. Orders slow and revenue growth fades.
Semiconductor manufacturing is among the most capital-intensive industries in the world and operating leverage works both ways; margins will compress faster than you can say “oversupply.” Small dips in utilisation can produce disproportionately large declines in profitability.
This creates an uncomfortable truth.
The buyers face rising costs as yesterday’s CAPEX becomes tomorrow’s depreciation. At the same time, the suppliers face slowing sales. Both experience margin pressure and a drop in earnings concurrently.
During the AI gold rush, almost everyone appeared to be winning.
As the capital cycle matures, there may be losers at both ends of the value chain.
That has important implications for investors.
Today’s financial statements capture the build-out phase of the cycle. They are a snapshot in time, not a steady state. Valuing AI businesses solely on today’s reported earnings risks extrapolating economics that may prove temporary.
Investors should focus less on where reported profits stand today and more on where the underlying economics are heading.
There is another consequence of the AI investment boom that deserves far more attention.
Consider how this extraordinary capital expenditure is being financed.
Until recently, the hyperscalers were among the most cash-generative businesses in the world, allowing them to fund growth almost entirely from internally generated cash flows. AI has begun to change that.
Capital-light businesses are becoming capital intensive.
Much of today’s investment is now being financed with debt, and that has important implications for shareholders.
The first is a duration mismatch.
If the economically useful lives of the assets being acquired is five or six years, while the debt used to finance it matures over ten or fifteen years, equity holders bear far more risk than creditors.
Debt holders have fixed contractual claims. Shareholders receive whatever remains. Should returns disappoint, enterprise value transfers away from equity holders long before debt investors begin suffering losses.
The second implication is more subtle.
For years, the hyperscalers have offset generous employee share-based compensation through substantial share repurchases, limiting dilution and supporting earnings per share. That too appears to be changing.
In the first quarter of 2026, breaking from convention, neither Alphabet nor Meta repurchased a single share. All available capital is now being redirected towards AI infrastructure spending.
Investors ought to be very concerned about this. If margins are squeezed and earnings decline while the sharecount expands, what does this mean for earnings per share and long-term shareholder returns.
The AI gold-rush is changing the financial characteristics of these companies as investments.
Acceptable Utilisation Rates
Everything discussed so far depends upon one critical assumption: the infrastructure eventually earns an acceptable return. This depends on whether or not the capacity being built out is properly utilised.
As discussed earlier, unused railway lines or fibre optic cable can remain productive for decades while demand gradually catches up with supply. AI hardware, on the other hand, is subject to the pace of technological progress, which rapidly erodes its competitive value.
In this scenario, time becomes the enemy.
AI infrastructure only creates value if businesses adopt it quickly enough to keep hundreds of billions of dollars of assets fully employed before technological obsolescence catches up. Ultimately, the AI investment thesis isn’t really about chip suppliers or hyperscalers; it’s about utilisation rates.
This shifts the focus.
Consumers ask large language models billions of questions every day. Businesses continue embedding AI into products, software development, customer support and internal workflows. None of that is in dispute.
The more important question is whether today’s demand represents durable economic value.
Uber recently disclosed that it exhausted its entire AI budget for 2026 within only four months. At that pace, annual spending would be roughly three times management’s original expectation.
Yet high consumption isn’t the same as high economic value.
More revealing than the spending itself was an admission by Uber’s CFO that there was a great deal of uncertainty over whether those costs were generating an adequate return. That raises questions about how sustainable current usage ultimately proves to be.
History is full of technologies that experienced explosive early adoption before settling into a much lower level of economically rational demand.
The same pattern may already be emerging inside large organisations.
Recently, employees of enterprises have been encouraged to maximise AI usage. Internal leaderboards measure token consumption. Teams compete to integrate AI into everyday workflows. The phrase “tokenmaxxing” has even entered the corporate vocabulary.
There is also anecdotal evidence that some frontier labs have heavily subsidised early enterprise adoption. Salesforce, for example, has reportedly benefited from a temporary ‘fee holiday’ on token consumption. If there is no cost, why not tokenmax?
That behaviour makes sense during the discovery phase.
Businesses want employees experimenting with tools that may eventually transform productivity. This is, after all, the ‘age of AI discovery’. But there’s an important distinction between using AI because it creates measurable economic value and using AI simply because it’s new.
The important question isn’t whether token subsidies exist. It’s whether demand remains equally strong once customers begin paying the full economic cost for artificial intelligence.
Eventually, experimentation gives way to measurement. Management begins asking a different question: “Does this spending generate an acceptable return?”
Economic value is ultimately determined by the customers’ willingness to pay for it.
That distinction will determine whether today’s infrastructure remains highly utilised or whether significant portions become economically redundant.
Microsoft has cancelled its Claude Code licenses only months after adopting it. The reason isn’t strategy, but the bill.
Other organisations are shifting workloads from expensive frontier models towards smaller, cheaper alternatives that produce sufficiently similar outcomes.
The objective is straightforward.
Reduce cost while preserving utility.
Perhaps this is a first clue to answering the utilisation question. The market is watching for other clues, in particular, for evidence that AI meaningfully raises profit margins outside the technology sector to justify the consumption expense.
Until that happens, much of today’s demand remains inadequate to underwrite an AI investment thesis.
This brings us to the final question: Who ultimately bears the cost of this extraordinarily expensive AI build-out?
The Ripple Effects Across the Economy
Semiconductor and memory chip manufacturers are currently enjoying exceptional margins because demand far exceeds supply for mission critical components.
When demand materially exceeds supply, prices rise. That’s how markets allocate scarce resources.
Bottlenecks capture huge economic rents, but that cost flows downstream.
Initially it may appear to be a cost born by the hyperscalers and AI model developers, but they aren’t spending hundreds of billions of dollars altruistically. They’re investing because they believe end users will eventually pay enough to cover the infrastructure bill, plus a profit margin.
That’s a significant leap of faith.
History suggests consumers rarely respond to new technologies by permanently increasing overall spending. Capital is, after all, a finite resource. So instead, they reallocate existing budgets.
That turns AI into a question of opportunity cost.
Every dollar spent on AI consumption is a dollar that can’t be spent elsewhere.
What does this mean?
Spending on software, equipment, hiring, advertising or other corporate priorities may have to give way. Investors therefore need to ask not simply whether AI changes the world, but where the money ultimately comes from.
The consequences extend well beyond the AI sector. The build-out has become large enough to influence prices across the wider economy.
Higher semiconductor and memory prices impact consumer electronics, industrial machinery, vehicles, medical devices and countless other products that rely on these components. Almost every manufacturer now competes, directly or indirectly, with AI infrastructure for scarce supply.
The explosive demand for AI infrastructure has caused supply constraints in all types of memory (DRAM, HBM, NAND, HDD).
Apple, despite largely sitting out the AI infrastructure land-grab, has already announced hardware price increases as component costs have risen. It won’t be the last.
There is no opt out. Whether companies invest heavily in AI or not, many are bearing its costs. Higher input costs are inflationary.
Apple appears to be sufficiently concerned. It has hedged itself by locking in $30 billion of Broadcom silicon supply through 2031, its biggest U.S. manufacturing deal ever.
Morgan Stanley estimates the impact of what it calls ‘chipflation’ on net margins by sector:
Some businesses will absorb those higher costs through lower margins. Others will pass them on through higher prices. Either way, someone pays.
If these cost pressures become broad enough, they eventually find their way into inflation. Central banks have little choice but to respond.
Higher inflation typically brings higher interest rates, raising the cost of capital across the economy and placing further pressure on investment, corporate profitability and valuations.
The irony is difficult to ignore. AI promises to lift productivity and lower costs over the long term, yet during the build-out phase it may temporarily produce the opposite outcome. Businesses face rising input costs while households contend with higher prices, long before many of AI’s productivity gains are widely realised.
If AI simultaneously displaces workers faster than productivity improvements emerge, economic growth could slow even as inflation remains elevated. Tighter monetary policy alongside slowing growth and rising ununemployment (when AI displaces people), are the defining characteristic of stagflation, which has historically resulted in falling equity valuations and periods of financial instability.
It’s a combination that can be particularly challenging for richly valued assets, especially those priced on expectations of strong future earnings growth.
This isn’t a prediction. It’s a risk that investors should consider when evaluating a market that appears to have priced AI almost exclusively through the lens of future benefits with little or no regard to near-term economic frictions.
The inflationary risk narrative doesn’t end there.
AI data centres consume extraordinary amounts of power, yet electricity supply cannot expand overnight. Building new baseload generation takes years, while ageing transmission networks have suffered from decades of underinvestment. Even where additional generation exists, moving that power to where it’s needed has become a growing constraint.
The economics are straightforward. When demand rises faster than supply, power prices increase.
Now think about electricity consumption. Manufacturers, retailers, hospitals, banks, offices and households all depend on it. Rising energy costs therefore spread through the economy in much the same way as higher semiconductor prices.
The benefits of AI may eventually arrive, but many of the costs are already here.
Financial markets have largely priced AI on the assumption that productivity gains will emerge rapidly across the economy. Yet many of the inflationary costs associated with building that future may lead to an economic squeeze on productivity, at least in the near term.
The irony is striking. If margins contract, then there’s less disposable capital for companies to allocate to AI consumption.
The problem is that adoption takes time. Healthcare operates within strict regulatory frameworks. Banks face governance and compliance requirements. Manufacturers must redesign production processes. Utilities, defence contractors, pharmaceutical companies and governments all move at their own tortoise pace.
Even where AI can deliver substantial productivity gains, organisations still need years to redesign workflows, retrain employees and integrate new systems.
That matters enormously for investors.
It creates a dangerous divergence between aggressive, front-loaded valuations today, and a much slower cash flow reality. Equity markets priced for instant earnings growth will face a painful repricing if the productivity hockey-stick takes six years rather than six months to arrive.
The technology can still succeed. The investment can still generate attractive long-term returns. But timing, once again, may determine who ultimately earns them.
Conclusion
While the long-term direction of AI appears increasingly clear, the investment thesis may not be. The timing and distribution of economic returns are anything but certain.
Throughout this article, we’ve challenged some of the assumptions underpinning the AI investment boom by asking the questions many investors seem reluctant to ask. Not because AI will fail, but because great technologies do not automatically produce great investments.
Today’s AI leaders appear priced for near perfection, leaving little margin for disappointment should reality unfold more slowly than markets currently expect.
None of this implies that the investment boom is irrational, nor does it suggest a collapse is inevitable. It simply reminds us that accounting, capital allocation and economics rarely move in perfect synchronisation. Timing matters.
History offers plenty of examples. Railways, radio, airlines, telecommunications and the internet all transformed society. They also experienced periods where capital flowed into the industry much faster than economic returns flowed back to investors, leaving years of disappointing performance before the underlying economics finally caught up.
AI is unlikely to be any different.
The technology may exceed our expectations.
The investment returns may not.
Economic evolution is another important consideration.
Commercial progress is shaped by survival of the fittest. For every Dell, Hewlett Packard and Apple, there’s a graveyard full of personal computer companies that failed despite operating in one of history’s greatest technology booms.
The AI industry is unlikely to be any different.
It’s a game of musical chairs. very company is investing as though it will still have a seat when the music stops. There won’t be. Some of today’s colossal investments will earn exceptional returns, but many won’t.
The trick isn’t knowing that the music will stop, it’s judging when to place bets on the winners. For now, the floor is packed, everyone’s dancing and optimism fills the room. But the clock on the wall is ticking louder than anyone cares to admit.
Strong framework.
What feels underappreciated is how much of this comes down to sequencing. In most capital cycles, returns do not emerge in line with investment, and rarely accrue evenly across participants.
The risk is less that AI fails, and more that capital, accounting and cash flows become misaligned - making both optimism and scepticism look more certain than they really are.
This essay gives structure to what many sense: the numbers don't add up, and accounting hides it. The depreciation wave will hit all hyperscalers at the same time, leaving no diversification for index investors. The $800B revenue math is likely understated, since AI's compute costs break software's high-margin model. And history agrees: the internet changed everything, yet Cisco still trades below its 2000 peak. Subsidized demand always rationalizes once real prices arrive, and Uber's blown budget is what the top of that curve looks like. Musical chairs, indeed.