A weekend read for Beaufort members
“Financial bubbles are then a phenomenon leaving a very complex legacy … after the bubble, the real infrastructural networks remain … and they generally achieve enough coverage during the frenzy to become positive externalities.” —Carlota Pérez
We know the inbox has been busy with AI. This note isn’t another opportunity or a teaser for one. It’s a step back: a weekend overview of what’s really going on, why we’re interested, where the risks sit, and how to think about AI in a portfolio without the noise.
The AI boom looks and feels like a bubble. The biggest tech firms on earth are spending like it’s wartime—building city-sized data centres, racing for chip supply, and hiring talent at a clip. The hype is breathless, the returns uncertain. As investors, the word “bubble” doesn’t exactly calm the nerves. Bubbles burst.
But history also suggests that periods of over-investment often lay the tracks for what comes next. The railways, the telecoms build-out, the dot-com era—each saw capital flood in, valuations run ahead of reality, and many companies fail. And yet, what survived beneath the rubble was infrastructure: the rails, the fibre, the cloud, and the talent base that made the next era possible. One generation’s bubble becomes the next generation’s broadband.
Today’s AI capex surge fits that pattern. Alphabet has guided to roughly $75–85 billion of spend in 2025, largely on AI servers, data centres, and networking. Meta has lifted its AI-related capex to roughly $64–72 billion, aiming to bring about a gigawatt of compute online this year and operate more than 1.3 million GPUs by year-end. Microsoft has committed around $80 billion in AI-focused capex in fiscal 2025, its largest allocation ever. Meta’s “Hyperion” cluster is designed to scale toward five gigawatts over several years—one site alone covering a footprint comparable to a sizeable chunk of Manhattan. This is industrial-scale building.
There’s a sober counterpoint: the bill. Data centres need chips, power, water, land, grid connections, and specialist staff. If current trajectories hold, global data-centre electricity demand could more than double by 2030, with AI as the fastest-growing driver. Most large models remain loss-leaders—expensive to train, expensive to run—underwritten by Big Tech balance sheets and venture capital. The economics must keep improving, or the maths won’t hold.
It’s not just companies; it’s countries. AI has become a national project. The U.S. is ahead, but China is pushing hard. Export controls limit access to the most advanced AI chips, and yet the strategic race continues. Nearly every cutting-edge chip still runs through Taiwan and TSMC’s fabs. As demand surges, the U.S. and Europe are trying to onshore production, but that is not simply a matter of building new facilities. The know-how is deep, granular, and hard-won over decades. A disruption around Taiwan would ripple through the global economy in ways that are hard to overstate.
Then there are jolts from unexpected places. DeepSeek, launched in early 2025, delivered GPT-4-level performance at a fraction of the cost and compute—reportedly around $6 million and roughly one-tenth the resources of comparable systems. For markets, it was a reminder that clever optimisation can challenge sheer spend. For everyone else, it hinted at a future where performance gains don’t only come from more money and more hardware.
If the corporate arms race feels unprecedented, the scale is not entirely new. The Manhattan Project cost about $2 billion at the time (roughly $38 billion today). That programme galvanised science and industry around an existential goal. The difference now is that the effort is being driven primarily by private companies with state-sized budgets. In the U.S., policy is pushing from behind—the Chips and Science Act, national R&D, and support for open-weight models—while firms like OpenAI, Anthropic, Google DeepMind, Meta and others compete at the frontier. The GPT-4o demo—real-time, multimodal, and fast—wasn’t just a product launch; it was a soft-power moment, the Silicon Valley version of a Sputnik beep.
And the money is flowing. Since 2023, CB Insights counts 73 $1B+ AI liquidity events—IPOs, M&A, reverse mergers or majority stakes—i.e., real cash-out moments for founders and investors.
What does all this mean for productivity and prosperity? The optimistic case is big: drug discovery and protein folding are accelerating; software development cycles are shortening; scientific research is being “co-piloted” by models that propose hypotheses, designs, and code. Some analysts put the long-run GDP uplift in the trillions. The more cautious case is that the benefits accrue unevenly, incumbent advantages harden, and the costs—from energy to data to compliance—prove stubborn for longer than expected.
So, are we in a bubble? Possibly. Are bubbles always and everywhere bad? Not necessarily. Pérez’s point is that the frenzy, while messy and painful, often leaves behind the infrastructure that powers the next wave. If this AI cycle were to deflate, we would likely still be left with denser global compute, smarter silicon, better networking, a workforce retooled on machine learning, and a clearer map of where AI truly adds value. The over-investment would have subsidised the future.
As investors, that doesn’t mean pile in. It means recognise what’s being built, be honest about the risks, and remember that price and timing matter. These are highly speculative bets where both risk and reward are amplified. Getting involved at the wrong moment can be disappointing—even if the long-term arc turns out to be transformative. The task is to separate the durable from the fashionable: chips and power versus buzzwords; workflows with measurable ROI versus demos; partnerships with distribution versus one-off pilots.
For the record, we’re enthusiastic about AI because, underneath the headlines, we see the pattern: heavy spending that looks irrational in the short term often builds the platforms that later feel inevitable. That doesn’t guarantee outcomes for any single company, and it certainly isn’t advice. It’s simply why we’re paying attention and why we’ve talked about it a lot.
“Private investments flowing into AI startups are creating billionaires faster than ever.” —Short Squeeze
Take this as a weekend read—a way to filter the noise on LinkedIn and elsewhere, and to consider where (and whether) AI might fit in your own portfolio over time. No pressure, no pitch. Just context, so you can make sense of what’s happening and move at your own pace.
Further reading
-
Carlota Pérez, Technological Revolutions and Financial Capital (book overview). Carlota PerezByrne Hobart & Tobias Huber, Boom: Bubbles and the End of Stagnation (Stripe Press). press.stripe.comBig Tech AI capex (Alphabet ~$85B ’25; Meta $64–72B; Microsoft ~$80B). Reuters+2Reuters+2IEA: Data-centre electricity demand could roughly double by 2030. IEAStanford HAI AI Index 2025: U.S. private AI investment $109B vs. China $9.3B (2024). Stanford HAIDeepSeek R1: low-cost Chinese model that shook markets. ReutersCB Insights: AI trends Q2 2025 (exits/M&A activity). cbinsights.com