NextFin News - The artificial intelligence boom is being funded by the largest corporate capital-spending cycle in history, and the question every investor is now asking is what happens when the bill comes due. Stephen Yiu, chief investment officer of the WS Blue Whale Growth Fund, has a clear answer: do not bet on the companies writing the cheques. Bet on the ones cashing them. His fund has exited Microsoft and Meta over AI return concerns and loaded up on the hardware layer - Nvidia, Broadcom, SK Hynix, Lam Research, Vertiv, Lumentum - that gets paid regardless of which AI application wins. The tension is whether that "pickaxe and shovel" logic really protects investors if the gold rush turns out to be smaller than advertised.
The Scale of the Bet Nobody Can Ignore
The spending is difficult to overstate. Amazon, Google, Meta and Microsoft plan to invest up to roughly $630 billion in capital expenditures in 2026, about 62% more than the $388 billion record set in 2025, with the overwhelming majority tied to AI data centres, GPUs and networking. Other trackers put the four-company total closer to $725 billion, a 77% jump year over year. Goldman Sachs estimates AI-related spending is tracking above $800 billion this year and sees $7.6 trillion flowing into AI infrastructure between 2026 and 2031, reaching $1.6 trillion of annual capex by the end of the decade.
This is not a normal investment cycle. It is a bet by four of the world's most valuable companies that the next wave of software will run on infrastructure they must build today. The risk, as Yiu frames it, is that enthusiasm for that promise is masking an earnings bubble in the companies funding it - the hyperscalers and the AI-application layer - while the suppliers of the physical plant are already collecting real revenue.
The market has not been blind to the tension. On 14 September 2026, leaders of the biggest AI companies - including Anthropic chief executive Dario Amodei, OpenAI's Sam Altman and Elon Musk - warned of potentially existential risks from the technology and called for slower advancement of frontier models. AI-linked stocks plunged worldwide on the news, shaking confidence in the industry whose vast infrastructure spending has driven global equity markets to record highs. A week later the Nasdaq notched a record-high close at 27,122.09 points, led by Advanced Micro Devices and AI stocks, with AMD reaching a $1 trillion market capitalisation for the first time. The S&P 500 closed about 0.4% below its 13 August record. In other words, the rally and the doubt are running side by side.
Yiu's own book reflects the split. As at 31 August 2026, the WS Blue Whale Growth Fund's top ten holdings were dominated by the infrastructure layer: Nvidia, Broadcom, SK Hynix, Lam Research, Lumentum, Vertiv, Modine, SanDisk and Applied Materials, alongside two deliberate diversifiers - Canadian Natural Resources in energy and a small position in London Stock Exchange Group. The top ten typically constitute more than half of the fund. Since its 2017 launch through June 2026, the fund returned 402%, versus 125% for the average IA Global fund.
The central question of this piece is simple: if the AI boom is real, the infrastructure names win. If the boom is a bubble, do they still win - or do they just fall more slowly?
The Pickaxe Thesis, and Why It Is Not a Free Lunch
Yiu's argument is a capital-cycle argument, not a technology forecast. In a June 2026 interview, the Blue Whale manager put the strategy directly:
"Instead of picking the AI winners of tomorrow, it makes more sense to position yourself within the AI infrastructure companies that are already receiving all the spending. They are already winning."
The logic has three parts. First, the hardware layer - GPUs, high-bandwidth memory, optical networking, liquid cooling - is fungible. On that point Yiu was blunt:
"If OpenAI just dropped out, whatever it has built in terms of physical infrastructure could be used by any other AI player. So that's not an overinvestment."
Second, the bottlenecks sit low in the stack, in hardware that is difficult to manufacture, supply-constrained and capital-intensive to build. Third, whoever wins the application race will still run on the same rails.
This is why the fund's largest positions map so directly onto the build-out. Applied Materials' tools modify materials at atomic scale and are used to produce virtually every new chip. Lam Research sits at the centre of the same equipment supply chain. SK Hynix holds a leading position in high-bandwidth memory, a product significantly more complex to manufacture and package than conventional memory that commands a meaningful premium. Lumentum makes the high-performance lasers that convert data into light as AI clusters outgrow copper interconnects. Vertiv provides the power and cooling infrastructure for data centres that require six to seven times the power of traditional facilities and relies on a network of almost 4,000 field service engineers. Broadcom, the fund's largest position, estimates that more than 99% of all internet traffic crosses at least one of its chips, and it builds custom silicon for customers such as Alphabet.
But the pickaxe thesis has a flaw that history exposes quickly. Pickaxes only hold their value if someone keeps digging. In the 19th-century railway mania, the companies that laid the track made fortunes - until the track was laid, capacity overshot demand, and a large share of railway capital was written off. The technology survived; many of the investors did not. The same pattern repeated in fibre optics during the dotcom boom: the world ended up with far more dark fibre than it needed for years, and suppliers that had priced in endless demand saw margins collapse.
Yiu is aware of this. His answer is that today's overinvestment is use-case-specific rather than thesis-breaking:
"At this moment in time, I would agree that money is being wasted, but that overinvestment is more of a use-case-specific issue."
He illustrated the point with a search-engine analogy: "Google Search was the seventh search engine created. The other six are gone. But it doesn't make online searching unreal." Dozens of social-media platforms competed, most died, capital was destroyed - and the market that emerged was larger than the sum of the companies that fought over it.
The padel-court analogy makes the mechanism concrete. A genuinely popular sport draws a rush of court construction; in some cities supply races ahead of demand; some courts sit empty; some developers lose money. The sport does not become unreal. The capital cycle simply sorts the winners from the losers, and "ultimately the winner of tomorrow will consolidate the entire market."
The question for investors is timing. Capital cycles do eventually clear - but "eventually" can mean years of dead money, and the suppliers at the top of the cycle are usually the last to feel the pain and the first to be re-rated when orders slow.
Cyclical or Structural: The Call That Decides Everything
This is where the analysis has to make a hard call, because the answer determines whether the infrastructure trade is a multi-year hold or a tactical position to be trimmed. The AI build-out is a structural shift wrapped in a cyclical capital boom - and the two layers point in opposite directions for investors.
The structural layer is real and durable. Accelerated computing has replaced traditional chip scaling as the engine of performance growth. Data centres are being rewired for optical interconnects because copper physically cannot carry high-speed signals across large AI clusters without power-hungry booster chips. High-bandwidth memory is now a design requirement, not an option, for training and fast inference. These are changes in how computing is built, and they will not reverse. Even if the AI application market disappoints, the infrastructure already deployed does not get un-deployed - it gets repurposed.
The cyclical layer is equally real and equally dangerous. Capital spending is surging because demand forecasts are extrapolating exponential growth. Goldman Sachs' own numbers show the velocity: the five largest U.S. hyperscalers - Amazon, Alphabet, Microsoft, Oracle and Meta - went from roughly $150 billion of annual spending in 2023 to an estimated $800 billion in 2026, and the bank now expects their AI infrastructure capex to rise more than 50% to $1.2 trillion in 2027, above the $1.1 trillion Street consensus. Its strategists estimate the group needs roughly $300 billion in annual AI revenue just to break even - a target current cloud growth has not reached. That gap between spending and monetisation is the definition of a cyclical overshoot.
The distinction matters because it tells you which parts of the trade survive a bust and which do not. A structural shift protects the companies with genuine bottlenecks and pricing power - the ones making HBM, advanced packaging, optical components and liquid cooling, where capacity is constrained and switching costs are high. A cyclical downturn punishes the companies that priced in perpetual exponential order growth and carry inventory or capacity that suddenly has no buyer.
This is why Yiu's book is not a simple "buy semiconductors" trade. It is a barbell: infrastructure names with supply constraints on one side, and deliberate non-AI diversifiers on the other. Canadian Natural Resources, with roughly 39 years of oil-sands reserves against about 10 for the average U.S. shale company and production with effectively zero decline, is a cash-flow asset with no AI exposure at all. The small London Stock Exchange Group position is a data-moat business that the market has recently de-rated as an AI loser.
The cyclical risk is that even good companies get re-rated when the order book slows. Lam Research and Applied Materials trade on equipment-cycle expectations; if hyperscalers cut capex, their earnings multiples compress even if their long-term technology position is intact. Vertiv and Modine depend on continued data-centre construction. SK Hynix's HBM premium depends on GPU demand staying ahead of memory supply. None of these are immune to a cyclical downdraft - they are simply better positioned within it than the companies writing the cheques.
The Second-Order Question Nobody Is Asking
The first-order effect of an AI bust is obvious: AI-related stocks fall. The second-order effect is more interesting, and it is where the real money is made or lost.
If hyperscalers cut AI capex, the immediate casualty is the equipment and component supply chain. But the second-order effect runs through the hyperscalers' own balance sheets and into the rest of the market. These companies have been spending $600 billion to $700 billion a year on infrastructure instead of buybacks, dividends and acquisitions. If that spending slows, three things happen at once.
First, free cash flow that was being absorbed by data centres gets released back to shareholders. The companies that look like losers in an AI boom - because they are not participating in the infrastructure gold rush - can become winners in an AI bust, as capital discipline returns. This is the mirror image of Yiu's "receiving end" thesis: in a bust, the payers keep the money.
Second, the capital that was crowding into AI infrastructure gets redeployed. Historically, capital cycles end not with the theme dying but with capital rotating to the next under-owned area - in the 2000s, out of telecoms equipment into emerging markets and commodities; in the 2010s, out of energy into technology. The beneficiaries of an AI bust are likely to be the sectors that today's AI allocation is starving.
Third, and most counter-intuitive, a bust could accelerate the very adoption it is feared to damage. When capital is cheap and abundant, every use case gets funded, including the bad ones. When capital is scarce, only the use cases with real economics survive - and the survivors inherit the infrastructure at depressed asset prices. Yiu's OpenAI example cuts both ways: if a leading lab fails, its infrastructure becomes available to competitors at a discount, lowering the barrier to entry for the next wave.
This is the transmission chain that matters: capex cut -> supplier earnings pressure -> hyperscaler free cash flow recovery -> capital rotation out of AI -> cheaper infrastructure for the survivors -> the next cycle begins on a healthier base. The bust does not end the theme; it clears it.
The Strongest Case Against the Thesis
The bear case against Yiu's positioning is not that AI is a fraud. It is that the market has already priced the "infrastructure is safe" consensus so completely that there is no margin of safety left.
Nvidia's shares trade at about 19.6 times expected 12-month earnings, a seven-year low and below the S&P 500's aggregate multiple of around 20 - but that cheapness is a function of earnings that have grown 65% in a year on $215.9 billion of fiscal 2026 revenue. If revenue growth decelerates sharply, the low multiple offers little protection: a company valued on growth loses its valuation support when growth slows, regardless of the absolute multiple. Analysts are anticipating $392.7 billion in Nvidia's current fiscal year - nearly double the 2026 figure. Any shortfall against that expectation is a re-rating event, not a buying opportunity.
The same logic applies across the supply chain. SK Hynix, Lumentum, Vertiv and Lam Research have all re-rated sharply on the expectation of sustained capex growth. Goldman Sachs' forecast of $1.2 trillion in hyperscaler AI capex in 2027 means the market is underwriting not just continued growth but accelerating growth. If the number comes in flat rather than up, these stocks can fall even though the long-term thesis is intact.
There is also a concentration risk that the pickaxe thesis does not fully address. The "fungible infrastructure" argument assumes many potential buyers for any asset that is built. In practice, AI infrastructure is highly customised - GPUs are tied to specific software stacks, custom silicon is built for a single customer, and data centres are located for power availability rather than market demand. If OpenAI "dropped out," its infrastructure would not be instantly usable by any competitor; it would require costly conversion, and some of it would simply be stranded. The dotcom-era fibre network was technically fungible too - but it sat in the wrong places, connected the wrong cities, and took years to absorb.
The most credible institutional version of this bear case comes from the other side of the same trade. JPMorgan strategist Dubravko Lakos-Bujas has argued that "2026 should be another strong year for AI stocks, with capex likely to surpass expectations," backed by a market that has returned 92% from its October 2022 low. Amundi chief investment officer Vincent Mortier put the timing problem more bluntly: "Whether there are excesses in the equity market on AI is no longer questionable, but to figure out which exact companies will be the losers and when this reckoning will happen is difficult."
The falsifying signal for the "infrastructure survives" thesis is specific and observable: if hyperscaler AI capex growth turns negative year over year for two consecutive quarters while AI revenue at the cloud platforms grows below 20% annually, the cyclical overshoot has begun and even the best-positioned suppliers will see multiple compression. At that point the trade is no longer "buy the bottlenecks" - it is "own the balance sheets that stop spending."
Who Benefits, Who Is Exposed, and What to Watch
The central judgment: the AI build-out is a structural shift in computing overlaid with a cyclical capital boom, and the two are being confused in the current debate. The structural layer protects the genuine bottlenecks - HBM, advanced packaging, optical interconnects, liquid cooling, semiconductor equipment - because the technology does not get un-invented. The cyclical layer will punish anyone who underwrites perpetual exponential order growth, including some of the very companies positioned as "safe."
The beneficiaries of a bust are, in order: the hyperscalers themselves, as freed-up capital flows back to shareholders and their own AI infrastructure becomes a sunk cost that starts generating returns; the supply-constrained bottleneck suppliers with pricing power, which survive the cycle and consolidate share; and the sectors starved of capital by the AI allocation, which become the rotation trade. The exposed are: the application layer with no revenue to justify its funding; the equipment and component names trading on continued capex acceleration rather than current earnings; and the custom-infrastructure builders whose assets cannot be repurposed.
Time horizons matter. In the short term - the next six to twelve months - sentiment and capex guidance drive prices, and a capex miss punishes the whole complex indiscriminately. In the medium term - one to three years - earnings separate the companies with real demand from those with slide-deck demand, and the survivors of the application shakeout begin to absorb the stranded infrastructure. In the long term - three years and beyond - the structural shift in computing asserts itself, and the companies that own the bottlenecks and the data moats compound from a lower base.
Scenarios. Base case: capex growth slows but stays positive, the application shakeout proceeds, and the infrastructure names consolidate at lower multiples before resuming growth. Upside case: AI revenue accelerates toward Goldman's $300 billion break-even threshold, validating the spending and lifting the whole complex. Downside case: hyperscaler capex turns negative for two consecutive quarters, triggering a broad re-rating across suppliers and a rotation out of the theme entirely.
What to watch. The falsifying signal is the one named above: two consecutive quarters of negative year-over-year hyperscaler AI capex growth combined with cloud AI revenue growth below 20%. Secondary signals: HBM pricing and lead times (a softening premium at SK Hynix signals demand cooling), optical-component order patterns at Lumentum, and data-centre power-and-cooling bookings at Vertiv and Modine. If those metrics hold up while application-layer companies fail, the pickaxe thesis is confirmed. If they soften together, the cycle has turned.
An AI bust would not prove the technology was a mirage - it would prove the capital cycle was a cycle. The survivors are not the companies that believed most fervently in the boom, but the ones that get paid while the believers argue about who wins.
Data as of 28 September 2026. This article is based on publicly available fund disclosures and manager interviews; it is not investment advice.
Explore more exclusive insights at nextfin.ai.
