A few companies are wildly overvalued. That doesn’t make AI a bubble.
On some afternoons this year, a single stock chart looked like a seismograph. Nvidia ticks down and pundits declare the end of the AI era. OpenAI’s valuation rumor ticks up and the bubble chorus returns. In the noise, a basic confusion keeps setting the agenda: we keep mistaking AI the technology and economy for AI the handful of model labs and a single dominant GPU supplier. The first is sprawling and early. The second is volatile and, in places, priced for perfection not for investment fundamentals. Those are not the same thing.
If you step back from the daily quotes, two realities stand out. First, the AI build‑out is broadening, not stalling. Second, when you do see froth, it’s concentrated in model providers and parts of the GPU supply chain, where rivalry, open‑weight competition, and falling prices compress margins. That pattern is a selective correction, not a macro bubble. (Gartner)
The confusion that powers the “bubble†narrative
A handful of names now carry an outsize share of market value and headlines. Regulators and central banks have warned about concentration risk and the chance of a sharp correction if sentiment on AI turns. That is a warning about how much is riding on a few leaders, not a claim that AI is worthless. In October 2025 the Bank of England’s Financial Policy Committee put it plainly: stretched, concentrated valuations raise the odds of a whiplash repricing. (Bank of England)
Concentration explains the spectacle. It doesn’t explain the economy.
Look where the money and momentum actually are
The AI economy is a stack: energy and real estate for data centers; core silicon and memory; servers and networking; cloud platforms and MLOps; models; and finally sector software and physical automation. You could see momentum at every layer in 2025:
Alphabet’s AI engine did not slow down. In Q3 2025 Alphabet crossed its first $100 billion quarter, guided $91–93 billion in capital expenditures for 2025 “to meet customer demand,†and said Google Cloud grew 34% year over year, led by AI infrastructure and generative AI solutions. That’s not hype; it’s booked revenue and planned build. (Q4 Capital Markets)
Microsoft’s cloud and AI flywheel kept accelerating. For the quarter ended September 30, 2025, Microsoft Cloud revenue grew 26%, Azure and other cloud services grew 40%, and commercial RPO jumped 51% to $392 billion. Satya Nadella framed it as “our planet‑scale cloud and AI factory†driving diffusion across Copilots and workloads. (Microsoft)
Google kept shipping new AI capability, not just models. This year brought TPU Ironwood disclosures and an expanded Vertex AI catalog at Cloud Next; new Agent Builder governance and observability for production agents; and even a privacy‑preserving Private AI Compute path to move heavier tasks off device with claimed safeguards. These are the plumbing and control planes enterprises actually use. (blog.google)
The industrial and energy bases are mobilizing. The IEA projects data‑center electricity consumption roughly doubling this decade in the base case, with AI a principal driver. That is forcing investment in power, cooling, and grid interconnects that sit far away from model demos. (IEA)
Robotics and edge are compounding. In 2024, factories installed 542,000 industrial robots, more than double a decade ago, and installations have held above 500,000 for four straight years. Edge and device AI keep spreading through PCs, phones, and industry endpoints, shifting where inference happens. (IFR International Federation of Robotics)
Enterprise adoption is broadening. The Stanford AI Index reports 78% of organizations used AI in 2024, up from 55% in 2023. U.S. private AI investment hit $109.1 billion in 2024, with generative AI drawing $33.9 billion globally. That is diffusion at scale, not a two‑company story. (Stanford HAI)
If AI were “just Nvidia and OpenAI,†these numbers and product roadmaps across Alphabet and Microsoft would be hard to square with reality. They square easily once you accept AI as a many‑market build‑out.
The model‑maker squeeze: when excellence meets economics
Why do headlines still read “AI bubble� Because the most visible part of the stack is also the most economically contested.
Quality convergence collapses moats. The 2025 Stanford AI Index documents how open‑weight models closed most of the gap with closed systems on popular leaderboards, narrowing from roughly 8% in early 2024 to about 1.7% by February 2025. When capabilities converge and switching costs fall, prices drop and margins compress. (Stanford HAI)
Price wars are real. In February, China’s DeepSeek announced aggressive off‑peak discounts of up to 75% for its APIs, explicitly pressuring rivals on cost. Even if you never use DeepSeek, the signal reverberates: inference gets cheaper, and model‑only business models get squeezed. (Reuters)
Meanwhile, customers keep using AI where value is measured, not just marveled at. Field studies show double‑digit productivity gains for customer support and software development teams using assistants and copilots. Less‑experienced workers benefit most, suggesting diffusion rather than displacement. That’s the profit engine enterprises care about. (NBER)
Put those three together and you get a tidy explanation: if there’s a bubble, it’s inside the model tier, not across AI as a technology. Price competition and open‑weight parity make a high‑margin, model‑only future unlikely for all but a very few.
Google and Microsoft prove the bigger point
My intuition is “there’s no money in making generic models; there’s money in using AI†and that is exactly what platforms are operationalizing.
Alphabet: beyond Gemini branding, the growth is in how customers use AI. Alphabet points to AI Overviews, AI Mode in Search, Gemini processing billions of tokens per minute via direct API, and a $155 billion Cloud backlog. It is also building the rails enterprises demand: confidential computing, security frameworks, and managed evaluators for agents. Those are leverage points, not loss leaders. (Q4 Capital Markets)
Microsoft: the firm’s quarter shows Azure +40%, and its narrative keeps returning to Copilots across high‑value domains. That is the monetization path you describe: embed AI into Office, Windows, Dynamics, GitHub, and industry solutions, then sell the outcomes. (Microsoft)
In both cases, the platforms are buyers, builders, and sellers of AI. They train and host models, yes, but the value capture is in distribution and integration: putting AI into the workflows, the operating systems, the developer tools, the call centers, the ad stacks, and the agent orchestration layers. That is much harder to commoditize.
What a selective correction looks like (and why it isn’t a pop)
Corrections have already hit parts of the stack that were priced for flawless execution. Super Micro Computer sold off when shipments and margins wobbled, even as AI demand stayed large; C3.ai retrenched on guidance. Those are firm‑level resets driven by mix, supply timing, and unit economics, not a verdict on AI writ large.
At the same time, core infrastructure revenue is still printing at scale, and hyperscaler capex is rising, not falling. A bubble bursts when demand evaporates. A correction occurs when expectations meet the limits of execution and competition. We are seeing the latter. (Q4 Capital Markets)
The rest of the iceberg: memory, power, networks, and robots
Even if you shaved multiples on the two celebrity tiers tomorrow, the rest of the AI economy would still be busy:
HBM memory has become a profit pool of its own. Analysts and trackers have SK hynix leading in HBM through 2025, with Samsung and Micron scaling hard and HBM4 on the horizon. That rivalry spreads value capture beyond any one GPU brand. (Counterpoint Research)
Power and cooling are now strategic constraints, which is why the IEA expects data‑center electricity demand to surge through 2030. That pulls in utilities, grid operators, and specialized thermal vendors. AI’s “picks and shovels†are literal. (IEA)
Robotics keeps marching. IFR reports 542,000 industrial robots installed in 2024, with Asia taking roughly three‑quarters of new deployments. Embodied AI is not waiting on a model leaderboard. (IFR International Federation of Robotics)
When people say “AI is just Nvidia,†they miss these engines of spend and revenue.
So, are we in an AI bubble?
No. We are in an AI build‑out with a model‑maker bubble risk at the frontier and a valuation concentration problem in a few tickers. The distinction matters. The stack is bigger than two logos; the customers are measuring value; and the infrastructure is still being poured.
If the last few quarters teach anything, it is that AI’s economics look healthiest when AI shows up in the product. The winners are building distribution, workflow fit, security, reliability, and control. The ones pricing access to generic tokens fight a race to the bottom.
Sources and further reading
Alphabet Q3 2025: revenue $102.3B, Cloud +34%, $91–93B 2025 capex; commentary on Search, Gemini, Cloud backlog. (Q4 Capital Markets)
Microsoft FY26 Q1 (quarter ended Sep 30, 2025): Azure +40%, Microsoft Cloud +26%, RPO +51%. (Microsoft)
Google Cloud and AI updates: Cloud Next 2025 highlights; Agent Builder observability and governance; Vertex AI release notes; privacy‑preserving Private AI Compute coverage. (blog.google)
Adoption and investment: Stanford AI Index 2025 on organizational use (78%), U.S. private AI investment $109.1B, and open‑weight catch‑up. (Stanford HAI)
Energy and infrastructure: IEA “Energy and AI†analyses on data‑center electricity demand. (IEA)
Robotics: IFR World Robotics 2025 press materials. (IFR International Federation of Robotics)
Price competition: Reuters on DeepSeek off‑peak API cuts up to 75%. (Reuters)
Selective corrections: Reuters on Super Micro volatility; coverage of C3.ai guidance resets.
Concentration risk: BoE FPC record and Reuters summary on risk of a sharp correction if AI sentiment turns. (Bank of England)
