This is not investment advice. Always conduct your own research and speak to your IFA before making investment decisions.

The Turtle Problem

There’s an old story about a scientist lecturing on cosmology. When he finishes, an elderly woman stands up and declares that the world sits on the back of a giant turtle. “What does the turtle stand on?” the scientist asks. “You’re very clever,” she replies, “but it’s turtles all the way down.”

In September 2025, watching Cursor reach a $10 billion valuation while sending every dollar of revenue to Anthropic, who burns money paying Amazon, who burns billions building infrastructure, the parallel became inescapable. In the AI economy, it’s not turtles—it’s negative margins all the way down.

I’m often asked: “Are we in a bubble?” My answer is always the same: “Yes, we’re in a bubble, and yes, there are still incredible investment opportunities.” Both can be true simultaneously. Bubbles can inflate far longer than sceptics expect. And sometimes, what looks like a bubble turns out to be the early, messy birth of a new economic order.

Meet the Players

Before we trace the money, let’s identify who’s burning it.

The Hyperscalers: Microsoft ($3.8T market cap), Google ($3.0T), Amazon ($2.3T), Meta ($1.9T), and now crucially, Oracle ($806B). Combined market cap: $11.8 trillion. Combined 2025 AI infrastructure spending: $365 billion. Combined AI revenue: $25-$30 billion across the major hyperscalers—a fraction of their infrastructure investment.

The Frontier Labs: OpenAI ($500 billion valuation), Anthropic ($183 billion), xAI ($200 billion) and Google’s DeepMind. OpenAI burns $14 billion annually on $15 billion revenue. Anthropic, despite exploding from $1 billion to $4 billion revenue in six months, still loses $3 billion on $7 billion annually. xAI expects just $500 million revenue in 2025. Combined, they’re valued at nearly $900 billion (plus Deepmind – not broken out separately) while collectively losing billions.

The Open Source Disruptors: Meta’s Llama 4, France’s Mistral, and critically, China’s DeepSeek and Alibaba’s Qwen. In January, DeepSeek R1 triggered a $600 billion market wipeout by matching GPT–4 performance. By July, Alibaba’s Qwen3 matched Anthropic’s Claude at $0.22 per million tokens—1/68th of Anthropic’s $15 price. They’re giving away what others sell for billions.

The Application Layer: Companies like Cursor ($10 billion valuation), Perplexity ($20 billion), and Europe’s Loveable ($4 billion). These companies build consumer-facing products, sending most revenue straight to model providers.

The Money Flowing Upwards Too

But the money is not just flowing down the stack in losses, it’s flowing back up the stack again as investments. Nvidia commits $100 billion to OpenAI, who uses it to buy Nvidia chips. Oracle signs a $300 billion deal with OpenAI—betting 37% of its market cap on becoming the AWS of AI. Microsoft invested $13 billion in OpenAI. Amazon put $8 billion into Anthropic with Google investing another $4 billion (alongside their internal Deepmind investment). The layers are funding their customers to buy their own products. Cisco used to do something similar with their ecosystem back in the dotcom era. But the numbers are bigger now.

The Trouble with Negative Margins – Following a Dollar Through the Stack

Here’s what happens when a developer pays Cursor $30 for monthly AI coding assistance.

Cursor receives that $30 and immediately sends it all to Anthropic, making the company “deeply unprofitable” even before counting salaries and servers. At a $10 billion valuation, investors are betting this changes dramatically.

Anthropic receives that $30 but spends $43 on compute. With $7 billion revenue and $3 billion losses, they’re burning 43 cents per dollar received. Their $183 billion valuation assumes compute costs collapse.

The hyperscalers collectively receive that $43 in compute payments. But this only goes to fund a fraction of the current capex load. It’s estimated that the hyperscalers are doing $25-$30bn of revenue on $150–200bn of incremental AI-related capex. Even if you assume depreciation over 5–6 years (generous, given the pace of technological change), the numbers don’t quite add up unless AI revenue increases dramatically.

Meanwhile, the frontier labs like Anthropic are counting on the per unit compute cost to decline meaningfully.

At the bottom of the stack sits Nvidia – the only company to make money today. They sell GPUs with 72% gross margins and 56.5% net margins. Every dollar of loss above them is Nvidia’s gain.

Capex has now reached a level where the hyperscalers have started funding investment with debt. Oracle issued $18 billion in bonds, its total debt approaching $100 billion. Google’s cash balance declined for two quarters. Amazon’s free cash flow turned negative. They’re collectively spending $63 billion more than their combined free cash flow, betting their stock prices stay elevated to support the leverage. Or that revenue starts growing fast enough to sustain the continued capex flow.

Could This Really Work?: The Bull Case

For current valuations to make sense, here’s what needs to happen at each layer:

Nvidia ($4.3T market cap): Must maintain 56% margins while growing revenue a further 60%+ despite AMD, Groq, and Cerebras building competing inference chips. They need AI compute demand to keep exploding faster than supply. Possible? Maybe. Physical AI is opening a new frontier. But competitive threats look daunting.

Hyperscalers ($11.8T combined): Need AI revenue to grow from $25 billion to $200+ billion within 3 years to justify current CapEx. Microsoft’s AI business already hit $13 billion run rate, growing 175% YoY. This far outstrips enterprise cloud growth rates. But $200bn+ is not impossible.

Frontier Labs ($900B combined): Either grow revenue from ~$25 billion to $100+ billion while maintaining pricing power against open source, OR pivot to application revenue with better margins. Plus ad monetisation. OpenAI’s ChatGPT shows a potential path to consumer revenue at high margins.

Applications ($50B+ combined): Must dramatically improve margins by switching to open source for commodity tasks, frontier models only for complex work, and raising prices as value becomes clear. If Cursor can charge $100+/developer/month with 50% margins using mixed models, this could work. Just about.

The Problem with Cheap Inference

But here’s the paradox: if inference costs collapse to pennies by 2027 as many predict, application companies win but infrastructure providers get crushed. If Claude costs $15 per million tokens today but falls to $0.50 by 2027, Anthropic’s revenue would need to grow 30x just to stay flat. Open source will commoditise basic AI. Pricing power evaporates. The very success of cost reduction destroys the business model of those who enabled it.

This is exactly what happened in previous platform shifts…

History Doesn’t Repeat, But It Rhymes

Mainframe Era (1960s–1980s): IBM captured all value. Software was bundled free with hardware. IBM’s gross margins exceeded 60%. Then distributed computing arrived.

PC Era (1980s–2000s): Value shifted to software. Microsoft and Oracle became giants while IBM struggled. Compaq and Dell commoditised hardware. Software had 80%+ gross margins; hardware dropped to 20%.

Internet Boom (1995–2020): Initially, hardware boomed—Sun Microsystems, Cisco, EMC hit massive valuations. Cisco reached $500 billion market cap in 2000. Then the bubble burst. Value shifted to applications: Google, Amazon, Facebook. Cisco still hasn’t recovered its peak.

AI Era (2022-?): Nvidia dominates early, capturing 56% margins like IBM once did. But if history holds, value will shift to applications. The question is timing.

The Bear Case: Three Ways This Ends

The Open Source Avalanche: Alibaba’s Qwen3 already more or less matches Anthropic’s Claude at 1/68th the price. When enterprises realise they can run these models on their own infrastructure, Frontier Lab revenue collapses. Without frontier lab payments, hyperscaler AI revenue stalls. The stack unravels from the middle.

The Margin Compression: Inference costs fall 95% as promised. Great for adoption, devastating for revenue. If tokens cost pennies, how does anyone charge dollars? The entire stack reprices lower. This will shift trillions of dollars in market value.

The CapEx Revolt: By Q2 2026, hyperscalers will have spent $500+ billion cumulative on AI infrastructure. If AI revenue hasn’t exceeded $100 billion, CFOs revolt. CapEx guidance gets slashed 50%. Nvidia’s forward orders evaporate. The music stops.

The AI boom doesn’t end when growth slows, but when capital markets stop believing margins will turn positive. That could be next quarter if markets revolt. Could be 2027 if discipline holds. Or massive enterprise value might be unlocked so quickly that margins suddenly flip positive across the board. Just don’t bet the house on all of it happening at once.

Why I’m Still Optimistic

Despite everything I’ve just written, I remain bullish on AI. History teaches us something crucial about bubbles: picking the right companies matters more than timing the market perfectly.

If you bought Cisco in March 2000 at the peak, you’d still be down nearly 20% today—25 years later. But if you’d bought Amazon at the same moment, you’d be up 65 times. Both were “overvalued” internet companies. Both crashed when the bubble burst. One recovered and transformed commerce. The other, despite growing revenue from $12 billion to $57 billion, never regained its peak valuation.

The AI boom will have its Ciscos and its Amazons. Yes, we’re seeing negative margins all the way down. Yes, $17 trillion assumes economics that don’t exist today. But somewhere in this stack, companies are building genuinely transformative technology. The productivity gains are real. The applications are useful. The value creation is happening, even if the value capture isn’t sorted yet.

My bet? Painful consolidation, not catastrophic collapse. The hyperscalers survive but rationalise spending. Oracle either becomes the AWS of AI or loses $300 billion finding out. Nvidia faces margin compression but remains dominant. The frontier labs will become application companies in their own right or consolidate into the hyperscalers. Applications bifurcate into winners using mixed models profitably and losers dying on negative margins.

The bubble may deflate. Valuations may compress. But for those who identify the Amazons of AI—the companies solving real problems with sustainable advantages—there’s still tremendous upside ahead.

That’s the paradox of bubbles: they’re simultaneously destructive and creative. The key is knowing which turtle you’re standing on, and whether it’s one that will survive the shakeout.