
Tech • AI • Robotics • Game
⏱ ≈ 4 min read · 25-min video — 21 min saved
Buying into AI is not the same as buying the flashiest AI stock, because much of the sector’s value may accrue first to the companies supplying chips, cloud capacity and electricity rather than to model makers themselves.
The starting point is the S&P 500, often treated as a broad, defensive holding. Yet a JPMorgan tally cited in the analysis found that 41 AI-linked companies accounted for nearly half of the index’s value in early October. That means many investors already have substantial AI exposure through standard index funds, often without realizing it.
Companies such as Microsoft, Alphabet and Amazon sit one layer above model developers because they own the data centers and increasingly design their own chips. Anthropic has reportedly committed about $900 billion in compute and infrastructure spending, with roughly half due before 2030, making it heavily dependent on outside suppliers. Those suppliers also benefit directly as investors, with Google and Microsoft participating in a combined $26 billion share purchase tied to the flotation.
Nvidia still captures the richest economics in the stack, with a gross margin above 70% for the past three years. That has made it look less like a normal hardware vendor than a toll collector for the AI boom. But the stock has already fallen roughly one third from its 6 October peak, and the larger strategic risk is that customers are trying to reduce dependence on its chips.
Anthropic, Google, Meta, Microsoft and Amazon are all exploring other chip options. Anthropic now spreads workloads across four chip families and says an internal model helped port Claude to AMD chips in three days, versus a process that once took six to twelve months. Even if that claim applies only to part of the workload, it signals a determined effort across the industry to weaken Nvidia’s grip.
Another recent AI-related listing showed how dangerous first-week buying can be. Shares were sold at $135, opened at $150, closed the first day at $161, and reached $225 four days later before falling back near $140 within five months. Investors who bought at launch did little better than break even, while those who chased the early spike suffered sharp losses.
At $83 a share, Anthropic would trade at roughly 18 times annualized revenue, implying a valuation near $1.5 trillion. The company has signed up more than half of the Fortune 500, but still represents less than 10 cents of spending for many large clients, and switching models can sometimes require only a single line of code. Its challenge is to become deeply embedded in workflows so customers cannot easily walk away.
Anthropic does not own most of the machines it relies on, while its infrastructure obligations are long-term and fixed. Its revenue, by contrast, remains usage-based and can slow quickly if clients cut demand or switch to rivals. That mismatch explains why valuation matters so much: if the company’s technological edge narrows, much of the upside could shift back to landlords and chip suppliers.
Basic AI tasks such as summarizing text or answering routine questions are already becoming commoditized. The more defensible market is complex work such as scientific research, debugging advanced systems and drug discovery, where errors are costly and customers may still pay premium prices. Anthropic’s investment case depends far more on dominating this harder segment than on selling low-cost general assistance.
One argument for Anthropic is that closed US labs may now be harder to copy. The company disclosed that three Chinese labs had queried Claude more than 16 million times through 24,000 fake accounts, apparently to distill its outputs into rival systems. Since access controls tightened and benchmark tests changed, the gap between the best open model and the top closed model reportedly widened from 3 points to 12, suggesting frontier labs may retain their lead longer.
The company is also trying to capture more value from discoveries made with its models rather than just charging for compute. It has built a biology unit, acquired Myril Medicine, and is pursuing two drug programs for rare muscular dystrophies. If successful, that would shift it from selling intelligence as a service to owning part of the downstream product, a far more lucrative but also more politically and regulatorily sensitive model.
A cautious framework places 70% to 80% of an equity portfolio in the broad index, 10% to 20% across major infrastructure and chip names, and no more than 5% in a single speculative company. On a €50,000 portfolio, that would imply roughly €40,000 in the index, €8,500 across mid-layer AI exposures, and about €1,500 in Anthropic, potentially split over time rather than committed in the first trading days.
The central question is not whether AI matters, but where its profits will settle across the stack. For now, Anthropic offers huge upside only if it preserves a technological lead long enough to become indispensable and capture more of the value created by its models.
Ask a question