A Long Story on AI. | Oct ‘25

It’s been a while since I last wrote, so I thought my newest Tab should be a comprehensive, bottom-up analysis of the AI rally, not only to offer some well-needed context, but also to reconcile the opposing strains of euphoria and scepticism that surround it. This will be a long read, but I hope it’s informative and compelling enough for you to form your own view on AI. I’ll focus mainly on three aspects: how AI is financed, the competing perspectives on its future direction, and the role of investor psychology in sustaining its seemingly endless rally.

1. AI Investment

Like any emerging technology (Personal computing, the Internet, social media, to name a few), AI’s early capital comes from the most optimistic, risk seeking investors: venture capitalists. These are funds that provide early stage funding for “high risk, high reward” ventures, especially those helmed by charismatic founders with good stories to tell. They invest in exchange for equity rather than immediate returns (think shark tank), which, like any other form of equity investment, meant their upside was uncapped.

Because venture capital bets on future viability, near-term profitability is often irrelevant. In the transactional world of finance, exit strategies are plenty, and as long as hype endures, liquidity follows. VCs can simply sell their stakes to the next eager buyer.

In just the first nine months of 2025, AI startups raised over $193B from VC funds — that’s a figure almost double from the year before in 2024. Yet, according to Andreessen Horowitz, venture capital firm, more than 80 per cent of that money went toward covering compute costs. Sam Altman himself has admitted that OpenAI makes a loss even on its priciest $200/mo ChatGPT Pro Plan, thanks to heavy usage by power users.

So, we’ve established the first link of the AI capital roadmap — VC → Startups.

From there, money flows to hyperscalers. These are massive data centre operators from which AI startups like OpenAI rent their compute capacity. Think companies like Amazon with AWS, Microsoft with Azure and Alphabet with Google Cloud. In the frenzy of the AI build-out, these companies have bled cash into infrastructure, not only to train their own models, but also to rent out to cash-strapped startups who cannot afford their own data centres.

Admittedly, these companies are sitting on bountiful cash reserves, but the picture isn’t so rosy when you scrutinise their balance sheet. When companies spend on costly infrastructure, they do not pay upfront. Instead, they spread the costs over the asset’s useful life through an accounting method known as depreciation, reflected as an ongoing Cost of Goods Sold (COGS) or operating expense on their income statements, depending on the asset's use.

Yet, a pressing set of questions concern the longevity of those fancy AI chips in their data centres. Last year, Nvidia announced it would unveil a fresh AI chip every year rather than every couple of years. In March, Jensen Huang remarked that “when Blackwell starts shipping in volume, you couldn’t give Hoppers away,” That would imply the useful life of Nvidia’s pricey chips is now a mere 12 months. Ironically, these hyperscalers have in recent times been raising their servers’ lifetimes to reduce depreciation charges in their accounts: Microsoft increased it from four to six years in 2022, Alphabet in 2023, Amazon and Oracle changed it from five to six in 2024. By raising server lifetimes, hyperscalers are able to spread depreciation charges over a greater time period, making their immediate profits look healthier than they really are. Like dishonesty in personal relationships, accounting trickery in finance is a fairly huge red flag.

That’s Startups → Hyperscalers sorted.

Where do Hyperscalers incur their costs? Quite obviously, with infrastructure manufacturers. The big names here are Broadcom, AMD, and TSMC. The biggest name, however, is Nvidia, with over 90% of the AI GPU market. Sitting at the top of the AI house of cards, it’s fairly natural to feel unsettled by the sustainability of continued capital flows from VCs. This, exacerbated by investor greed, has led Nvidia to make some questionable investments.

Summary of Capital Flows in AI

VC → Startups → Hyperscalers → Nvidia (mostly)

2. Creating value out of thin air

Recently, word got out that AI might be in a bubble phase, and from some fairly influential people at that. Sam Altman, Jeff Bezos, and David Solomon all share a bearish take. Being the golden goose of the AI sphere, investors look to Nvidia with seemingly unrealistic expectations. The outlook isn’t for a 30% increase in revenue year-on-year, but a 100% one.

While VCs have been making up the difference for the astonishing rise in demand for AI chips, fears have emerged that the sky may not actually be the limit for the amount of money VCs are willing to flush down the toilet. In other words Nvidia fears that VCs are nearing a liquidity crunch, and so are startups, and hyperscalers…

Nvidia’s solution? To take matters into its own hands by stimulating demand with its own cash.

The headline deal is its progressive $100B investment in OpenAI to expand its compute capacity consumption with Hyperscalers, which, unironically, run on Nvidia chips. Perhaps Nvidia thought it was a wise investment to own a stake in OpenAI, but at its latest $500B valuation for a mere $12.7B in annual revenue, OpenAI is valued at a price to sales of 39. That’s discounting the fact that OpenAI is unprofitable and has a negative cash flow. Capex and operational costs unconsidered, it’ll have to make hundreds of billions in the coming years for the investment to pay off.

Less conspicuous are Nvidia’s deals with Coreweave and Lambda — two emerging hyperscalers. In one arrangement, Nvidia committed $6.3 billion to purchase excess capacity from CoreWeave, effectively insuring it against a demand slum. It comes at a point where Coreweaves credit to expand is stretched thin.

To explain this with a relatable analogy: suppose you’re buying a car for Uber, but the bank refuses to extend credit because it doubts your ability to repay from Uber income alone. A desperate car dealer then offers to pay you occasionally to drive their employees, convincing the bank you have guaranteed income — so you buy five cars instead. The car dealership profits, you get your cars — everyone is happy.

But you can’t drive five cars.

On the other hand, the Lambda deal suggests that Nvidia is willing to pay $1.5B to lease their GPUs back, which is questionable, given Nvidia’s apparent lack of intense processing needs. Ultimately, one thing is clear: Nvidia is trying to fan optimism before people become disinterested in AI.  Naysayers will likely take it as a hint that the end is near.

3. History doesn’t repeat itself, but it often rhymes

In an earthquake, the ground level always seems to be the safest. Likewise, in highly uncertain technological revolutions, the safest bet will always be infrastructure firms. In a sector’s infancy, it’s often impossible to know which startups will make it out alive, but as long as the sector survives, as a majority of investors concur, infrastructure firms almost certainly will too. As it turns out, in both scenarios, that is not always the case.

Global Crossing was a dot-com giant that laid undersea internet cables. In the early 1990s as internet adoption accelerated, Global Crossing scaled immensely, borrowing billions to fund cable construction. Yet towards the final years, the anticipated demand had not materialised, and Global Crossing resorted to accounting manipulation to inflate revenue.

It did this through capacity swaps with other broadband companies. For instance, Global Crossing would sell $100M worth of capacity on one stretch of its cables for $100M worth of capacity on another company’s cables. This was a contractual agreement and no cash was actually exchanged. Nonetheless, Global Crossing recorded it as revenue, inflating its reported figures and sustaining its massive stock valuation. A few years later, Global Crossing imploded.

For cash-rich incumbents like Nvidia, an implosion is unlikely, but Coreweave? That’s another story. Again, history often rhymes.

4. But it’s not all good news!

And you’d be right. Around 25-30% of CNBC’s AI reporting errs on the side of caution, compared to the euphoric delusion during the internet boom. Even the shopping mall janitor could tell you that AI is in a bubble, and that certainly adds weight to the bearish views.

I contend that human fallibility prevails even in the face of an imminent disaster. Knowing is one thing; having the discipline to act on prevailing knowledge is quite another. Greed often endures. It would certainly be difficult to persuade someone to remain on the sidelines when Nvidia has risen 117% since April, and Google 86% over the same period.

FOMO and peer pressure will eventually draw people into the markets, regardless of their awareness of a bubble. Case in point: junk bonds now offer spreads of only 2.8% over government bonds, well below the ~4.8% average over the past two decades. Corporate bonds from companies such as Microsoft yield even less than government bonds.

But at its current state, the AI rally seems a lot more restrained than the internet era, not least because few truly AI-only companies exist. Google isn’t reinventing the wheel with AI, it’s integrating it into existing products which are proven and profitable. Meta isn’t selling AI, it’s bolstering its ad targeting business with it. Evidently, the sheer scale of these AI contenders also injects a surprising amount of conservatism in their AI strategy. Suppose one day Google decides to ditch its advertisement and search empire for selling AI chatbots, I think that’s when I would GTFO, and I recommend you to, too.

5. The Call to Action

Oftentimes, a rally has no clear signs of discontinuation until right before it implodes. The AI bubble could burst tomorrow, or in five years, or not at all. If you shorted the market today, you’d be out of the game in no time. If you wait it out, you’d incur substantial opportunity costs.

There’s no point in timing the market — but there is value in optimising your portfolio to capture the upside while minimising downside risk. That means investing judiciously in companies that aren’t putting all their eggs in the AI basket — firms with solid fundamentals that remain profitable even if AI turns out to be a dud. It also means prioritising companies with established avenues to implement AI, rather than those treating it as an experimental side hustle. And most importantly, using financial performance to cut through the storytelling.

I genuinely believe AI has the potential to be revolutionary — but so long as I still have to do my own laundry, that revolution isn’t here yet. After all, efficient markets mean survival of the fittest, not survival of the best con artist.