In partnership with
Dear Readers,
Today, the markets are booming: record profits, billion-dollar valuations, and investments at an unprecedented pace—and at the same time, the question of whether it's all too good to be true. The parallels to the dot-com era are striking, but this time it's not about cables in the ground, but data centers, GPUs, and gigantic amounts of energy. There is a fine line between euphoria and warning signs, and that is precisely where the debate is taking place: bubble or basis for a new industrial revolution?
In this issue, we take a look at the breathtaking capital cycles surrounding Nvidia, OpenAI & Co., analyze where exaggeration ends and substance begins, and show the hard indicators beyond the headlines: energy bottlenecks, real sales, long-term contracts. We also take a look into the future: What will happen when AGI really takes shape—and what upheavals we can expect in the labor market and society. Dive in—the answers will surprise you.
All the best,
A bubble about to burst?
The TLDR
The current AI boom exhibits a paradoxical mix of a speculative bubble and a real infrastructure revolution. While the "application layer" shows classic bubble signs—soaring valuations, record-breaking venture capital deals, and unprofitable startups—the underlying "infrastructure layer" is anchored in tangible, physically scarce resources. Massive, long-term investments in data centers, energy, and specialized chips, coupled with real revenue from major AI players and measurable productivity gains, suggest that while a shakeout of overvalued AI applications is likely, a complete dot-com-style crash is not. The future of the AI economy will be determined less by hype and more by the real-world constraints of energy, hardware, and sustainable business models.
Looking at the markets today, one sees a paradoxical simultaneity: record highs and rumblings of discontent. Nvidia climbs above $4 trillion in market capitalization, while analysts label the same charts with the word “bubble.” At the same time, the major platforms are investing at a pace that makes even the cloud decade seem modest. Alphabet raised its annual CapEx forecast to $85 billion in July, Microsoft reports declining cloud gross margins due to the expansion of AI infrastructure, and Dell'Oro expects global data center investments to break the $1.2 trillion mark by 2029. At the same time, the International Energy Agency estimates that data center electricity consumption will double by the end of the decade, driven by AI workloads.
On the other hand, capital is pouring into start-ups like never before since the dot-com years. OpenAI closed a round of up to $40 billion in 2025, and Anthropic added $13 billion shortly thereafter, valuing it at $183 billion. According to CB Insights, US$47.3 billion flowed into AI start-ups in the second quarter of 2025 alone, including 1,403 deals. If that sounds like the “late 1990s,” it's no coincidence.
The key question that arises from this is less simplistic than “bubble: yes or no?” What is interesting is whether we are seeing a sectoral bubble in the application and venture segment atop a substantial, physically anchored infrastructure cycle—and how likely sustainable returns are within this stratification.
Ad
Your boss will think you’re a genius
If you’re optimizing for growth, you need ecomm tactics that actually work. Not mushy strategies.
Go-to-Millions is the ecommerce growth newsletter from Ari Murray, packed with tactical insights, smart creative, and marketing that drives revenue.
Every issue is built for operators: clear, punchy, and grounded in what’s working, from product strategy to paid media to conversion lifts.
Subscribe for free and get your next growth unlock delivered weekly.
The signs of a bubble
Valuations are outpacing cash flows. Nvidia's market capitalization rose from $3 trillion to $4 trillion in just over a year; the picture is similar for private champions: Anthropic tripled its valuation within months. Such multiple expansions are reminiscent of phases in which expectations outpace fundamentals.
“AI company Anthropic zipped through its new fundraising, collecting $13 billion in a Series F round at a $183 billion post-money valuation. Anthropic’s valuation nearly tripled since March, when it was worth $61.5 billion.” (Wallstreet Journal)
CapEx rally with “overbuild risk.” Breakingviews pointedly stated that the AI boom is “infrastructure disguised as software.” McKinsey estimates that inference-oriented data centers alone will require at least $3.7 trillion in investment—with a corresponding appetite for electricity. Such front-loaded spending carries the classic risk of overbuilding: hardware becomes obsolete faster than it can be depreciated, utilization rates fall short of plans, and projects start with overly optimistic assumptions. The Financial Times (FT) recently drew parallels to the telecom overbuilding euphoria of the late 1990s.
“The pace of development has drawn comparisons with the telecoms bubble in the late 1990s, when companies laid more than 80mn miles of fibre optic cables across the US in a drastic overestimate of the demand required. The glut meant costs plummeted and many companies failed.
“People are making forecasts on the assumption that all enterprises will start to use AI technology and pay for it, and pay enough for it to justify the return on investment for all these training facilities,” said a banker who works on data centre deals.” (Financial Times)
Financing structure and interest rate channel. Part of the billions in funding required is not coming from the hyperscalers' operating cash flows, but is being outsourced to the periphery via private credit, project financing, and sale-leasebacks. In such leveraged chains, fluctuations in demand have a disproportionate impact—a well-known bubble pattern.
A start-up market with signs of a boom. $47.3 billion in Q2/25, record-breaking “megadeals,” aggressive deal pace—it's hot. Even if part of these rounds flows into the real economy in the form of computing power (GPUs, HBM, colos), the failure rates of later cohorts are statistically high.
Unit economics remain mixed. Microsoft explicitly points to strong cloud gross margins due to AI expansion—an indication that token prices and inference costs (energy, HBM bottlenecks) are limiting the scale story in the short term. When prices per unit of output fall faster than the cost per computing operation, monetization becomes a test of patience.
“Revenue increased $36.6 billion or 15% with growth across each of our segments. Intelligent Cloud revenue increased driven by Azure. Productivity and Business Processes revenue increased driven by Microsoft 365 Commercial cloud. More Personal Computing revenue increased driven by Gaming and Search and news advertising.Cost of revenue increased $13.7 billion or 19% driven by growth in Microsoft Cloud.Gross margin increased $22.9 billion or 13% with growth across each of our segments.•Gross margin percentage decreased slightly driven by Intelligent Cloud, offset in part by More Personal Computing.•Microsoft Cloud gross margin percentage decreased to 69% driven by the impact of scaling our AI infrastructure, offset in part by efficiency gains in Azure.Operating expenses increased $3.8 billion or 6% driven by investments in cloud and AI engineering and Gaming, including the impact of the Activision Blizzard acquisition.” (Microsoft)
The evidence against a bubble
Actual sales instead of pure “eyeballs.” OpenAI reported annualized sales in the low double-digit billion range in 2025; Anthropic speaks of a run rate of around $5 billion. This may be ambitious in relation to valuations, but it is real – unlike many dot-com business models. In addition, hyperscalers are reporting accelerated growth in cloud segments with AI components (Copilot, Vertex, Bedrock).
Shortage of physical bottlenecks rather than purely narrative scarcity. Electricity and grid connections are the hard currency. Data center vacancies in North America fell to ~1.6%, large power blocks >10 MW are becoming scarce, and grid connections are booked up for years. Such bottlenecks are atypical for pure financial bubbles.
Substantial, long-term energy contracts.
Microsoft, Amazon, and Google are signing 20-year PPAs—even going so far as to restart decommissioned nuclear power plants (Three Mile Island) or pursue SMR plans. Those who accuse them of “vibe spending” must explain why corporations are simultaneously signing fixed 24/7 contracts for nuclear energy.
“U.S. power demand from data centers is forecast to double within five years, rising from 176 TWh in 2023 to between 325 and 580 TWh in 2028, government data shows. Big Tech companies favour adjacent dedicated power generation to ensure a steady supply of energy around the clock. In October, Google signed an agreement with Kairos Power to develop 500 MW of SMRs close to data centers from 2030.” (Reuters)
“Picks and shovels” are earning money today. Memory is the new bottleneck: SK hynix and Micron are reporting record sales and margins thanks to HBM demand. This is fundamentally different from the 1990s, when suppliers often delivered too early and too cheaply for fiber optic expansion.
““It became clear decades ago that the commodity dynamic in the memory market would make it very hard to make outsized profits,” says Miller. That prompted many of the brightest minds, and ambitious entrepreneurs like Nvidia’s Jensen Huang and Qualcomm’s Irwin Jacobs, to turn their attention to processor chips, he adds. “But now, memory is back.” (Financial Times)
Productivity evidence beyond anecdotal evidence. In experimental studies, the OECD finds robust productivity gains of 5–25% in activities such as writing, coding, and consulting—today, mind you, not in the distant future. McKinsey sticks to estimates of a global annual value contribution in the multi-billion range. That's not “traffic,” that's output.
Dotcom comparison: What is reliable, what is misleading?
The most superficial comparison is the striking price chart. It is better to look at the structure. At the end of the 1990s, fiber optics were laid on a massive scale – much of it ended up stranded, but the networks formed the basis for Web 2.0, streaming, and mobile. Something similar is happening today: we are building a computing and energy base load for probabilistic software. In both cases, investors often confused timeline (too early) with direction (still correct). The difference: the big buyers (hyperscalers) are liquid, integrate vertically (own chips, own PPAs), and can cross-finance utilization. This reduces systemic risk—and shifts it to smaller, externally financed operators.
Interim Result
Scenarios, indicators, preliminary results
Scenarios. A correction is plausible in the short term—price declines of 20–40% in parts of the ecosystem are not the end of the world if utilization and cash flows follow suit. A selective “application shakeout” (many agent/thin wrapper start-ups, few platform winners) is likely from a venture perspective. Infrastructure winners (GPU/HBM/network/power oligopoly, colocation REITs) remain cyclical but structurally supported.
What to look out for? (i) Utilization of new capacity (>80% from year 2), (ii) AI revenue share in cloud segments (Azure/Google Cloud/AWS) and associated margin trends, (iii) price/cost path per token/query relative to energy and HBM costs, (iv) power allocation (waiting times, grid bottlenecks), (v) deal mix in venture (fewer mega rounds, more M&A/down rounds). The available data currently points to genuine scarcity and gradual margin leverage (efficiency gains), not just hot air.
Preliminary findings. The strongest bubble signals are found in the venture/application layer (valuations, deal pace). The infrastructure layer (chips, memory, construction, electricity), on the other hand, shows signs of a difficult, capital-intensive installation cycle – with real bottlenecks and reliable purchase agreements. It is unlikely that there will be a crash like in 2000, but rather a rotation: from narrative plays to cash flow generators; from training excesses to inference economics; from unprofitable growth to productivity-driven ROI in core processes.
Despite all concerns, research into generative AI is continuing with full force. And ultimately, there are no signs of a slowdown in the scientific sector.
Conclusion
The initial question can be answered in a nuanced way. Yes, there are bubble-like zones: late VC rounds, copycat apps, “AI for everything” without defensible moats. These segments will thin out—and that's healthy. At the same time, hard indicators argue against the big needle: real revenues from the models, scarce physical resources, long-term PPAs, record-low vacancies, and an ecosystem in which the picks and shovels are already generating profits today. Unlike in 2000, liquid anchor customers are financing a large part of the expansion themselves, which limits systemic risks even if individual projects fail.
More likely than a big bang is a multi-year, wave-like transition toward robust productivity utilization: first installation (CapEx, networks, electricity), then diffusion (use cases, processes, pricing). For investors, this means factoring in cyclical setbacks but taking the long-term real economy of computing and energy infrastructure seriously. Or, to quote Breakingviews: “The AI boom is infrastructure in software clothing” – which is precisely why it will not behave entirely like a software bubble.
Sources:
Sponsored By Vireel.com
Vireel is the easiest way to get thousands or even millions of eyeballs on your product. Generate 100's of ads from proven formulas in minutes. It’s like having an army of influencers in your pocket, starting at just $3 per viral video.
Chubby’s Opinion Corner
A bubble or not—but what happens when AGI arrives?
We are currently in a strange limbo – half industrial base load expansion, half speculative house of cards. Whether it is a “bubble” depends on which layer you look at.
The infrastructure layer (chips, memory, power, data centers) seems as real as a new energy sector: physical shortages, long delivery times, massive CapEx, 20-year contracts. That won't just disappear, even if some of today's startups implode. Similar to the fiber optic expansion of the 2000s: the companies disappeared, the cables remained – and later became the basis for streaming and the cloud.
The application layer, on the other hand, is full of bubble signals: unprofitable startups with billion-dollar valuations, copycat products, dizzying valuations. There will certainly be a shakeout here. Venture capital traditionally works according to the logic: 1 unicorn carries 9 write-offs. That feels painful from an investor's perspective, but it's not a systemic break.
Things will get exciting with the transition to AGI: if the models really do more than just generate code or write texts, but independently plan, develop strategies, and drive innovation, then the equation will be turned on its head. Labor markets could not only shift, but collapse: knowledge work, consulting, even research – sectors that were previously considered “safe” – would come under pressure. At the same time, there would be an enormous productivity boost, which in theory would generate prosperity, but in practice could exacerbate social upheaval (inequality, concentration of power, geopolitical tensions).
Whether this bursts or sustains a bubble depends less on “investor imagination” than on how quickly and how deeply real productivity from AI systems seeps into the economy. If AGI does indeed arrive in the 2030s, it will not be a bubble, but a paradigm shift – comparable not to dot-com, but to industrialization. However, the timing remains unclear, and that is precisely where the uncertainty lies.