The AI industry started 2025 acting like money had no ceiling. By the end of the year, it had something it hadn’t really faced before: a vibe check.
The cash is still flowing, the valuations are still wild, and the rhetoric about “world-changing” models hasn’t cooled much. But the mood has. Between trillion‑dollar ambitions, circular funding loops, regulatory pushback, stalled magic in model releases, and grim trust-and-safety stories, 2025 was the year AI stopped being celebrated uncritically and started getting interrogated.
The money got surreal — then awkward
Early 2025 was peak AI exuberance.
- OpenAI raised $40 billion in a SoftBank-led round at a $300 billion post‑money valuation. It’s also in talks to raise $100 billion at an $830 billion valuation, edging toward the $1 trillion IPO target it’s reportedly chasing for 2026.
- Anthropic pulled in $16.5 billion across two rounds this year, pushing its valuation to $183 billion, with investors like Iconiq Capital, Fidelity, and the Qatar Investment Authority. In a leaked memo, CEO Dario Amodei told staff he was “not thrilled” about taking money from dictatorial Gulf states — a rare bit of candor amid the funding frenzy.
- xAI, Elon Musk’s lab, raised at least $10 billion after acquiring X.
The froth wasn’t limited to the giants.
- Former OpenAI CTO Mira Murati’s stealth startup Thinking Machine Labs landed a $2 billion seed round at a $12 billion valuation before describing a real product.
- “Vibe‑coding” startup Lovable raised $200 million in Series A eight months after launch, becoming a unicorn. This month it added another $330 million at a nearly $7 billion valuation.
- AI recruiting startup Mercor pulled in $450 million across two rounds, reaching a $10 billion valuation.
All of this is happening while enterprise AI adoption is still modest and infrastructure is visibly straining. That disconnect is why fears of an AI bubble — and of who’s left holding the bag — got louder as the year wore on.
Build, baby, build — and hope the grid holds
Those stunning valuations aren’t just for flashy demos. They’re underwriting an unprecedented infrastructure binge.
Much of the capital being raised is effectively pre‑booked for chips, cloud contracts, and power. In some cases, the same investors or partners funding AI labs are also selling them compute, blurring the line between real demand and financial engineering. OpenAI’s infrastructure‑linked funding arrangements with Nvidia became the emblem of this circularity.
A few headline deals show the scale:
- Stargate, a joint venture between SoftBank, OpenAI, and Oracle, includes plans for up to $500 billion in AI infrastructure spending in the U.S.
- Alphabet acquired energy and data center infrastructure provider Intersect for $4.75 billion, just as it announced plans to lift compute spend in 2026 to $93 billion.
- Meta accelerated its data center expansion, pushing projected capex to $72 billion in 2025 as it races to secure enough compute to train and run next‑gen models.
Then the cracks started to show.
Blue Owl Capital, a private financing partner, walked away from a planned $10 billion Oracle data-center deal tied to OpenAI capacity. That pulled a thread on the assumption that infinite capital would simply materialize for every hyperscale build‑out.
Meanwhile, grid constraints, soaring construction and power costs, and growing community and political resistance are slowing projects on the ground. Senator Bernie Sanders, among others, has called for reining in data center expansion. Local fights over water usage, land, and energy mix are turning AI infrastructure into a frontline political issue.
The money is still enormous. But it’s running up against the physical realities of steel, land, and electrons.
The model “magic” faded
From 2023 through 2024, every big model release felt like a plot twist. GPT‑4, GPT‑4o, Gemini 1.5 — you could feel the step change.
2025 broke that streak.
GPT‑5 landed with expectations sky‑high. On paper, it was a clear improvement. In practice, it didn’t hit with the same shock as GPT‑4 or 4o. The pattern repeated across the space: better benchmarks, clever new tricks, but fewer moments where users felt they’d leapt into a new era.
Gemini 3 is a good example. It’s topping several benchmarks and is widely seen as a technical success. But its real significance is more mundane: it mostly brought Google back to parity with OpenAI, the level that prompted Sam Altman’s infamous internal “code red” memo and a scramble to defend OpenAI’s lead.
The more interesting story was where breakthrough models came from.
Chinese lab DeepSeek launched R1, a “reasoning” model that competed with OpenAI’s o1 on key benchmarks — and reportedly did it for a fraction of the cost. The message to investors and incumbents was blunt: frontier‑class models aren’t the exclusive domain of a handful of U.S. giants anymore, and they don’t necessarily require tens of billions in spend.
Once the leaps between model generations get smaller, the conversation shifts. And it did.
From big models to actual business models
By mid‑2025, investors and customers were asking a different question: Who can turn all this capability into a product people actually rely on and pay for?
That pressure showed up in some aggressive experiments.
- AI search startup Perplexity briefly floated a plan to track users’ browsing to sell hyper‑personalized ads.
- OpenAI reportedly considered charging up to $20,000 per month for highly specialized AI offerings — a high‑stakes test of how much enterprises will pay for cutting‑edge models.
The real battlefield, though, is distribution.
Perplexity is trying to own more of the user surface area with its Comet browser, pitched with agentic capabilities. It also struck a $400 million deal with Snap to power search inside Snapchat, effectively buying access to an existing firehose of users.
OpenAI is running a parallel playbook: turning ChatGPT from a chatbot into a platform.
- It launched the Atlas browser and consumer‑facing features like Pulse.
- It started letting developers and companies ship apps inside ChatGPT.
- It’s amping up its enterprise sales motion, pitching ChatGPT as the hub where workers spend a good chunk of their day.
Google is leaning into its incumbency.
- On the consumer side, Gemini is being baked directly into Google Calendar and other everyday products.
- On the enterprise side, Google is hosting MCP connectors to make its cloud and productivity ecosystem harder to rip out.
When new model launches feel more incremental than revolutionary, owning the user relationship — and the billing relationship — becomes the real moat.
Trust, safety, and the dark side of “engagement”
If 2023–2024 were about AI’s economic promise, 2025 forced the industry to confront its liabilities.
More than 50 copyright lawsuits against AI companies rolled through courts worldwide. At the same time, reports of so‑called “AI psychosis” — cases where chatbots allegedly reinforced delusions or contributed to suicides and life‑threatening behavior — turned AI trust and safety into a public health issue, not just a PR one.
On copyrights:
- Anthropic agreed to a $1.5 billion settlement with authors, closing one high‑profile front in the battle over how training data is sourced.
- The broader fight is shifting from “you can’t use copyrighted content” to “you can, but you must pay for it.”
- The New York Times is suing Perplexity for copyright infringement, a case watched closely as a potential template for future compensation regimes.
On mental health and safety:
- Multiple reported deaths by suicide and severe delusional episodes followed prolonged interaction with AI chatbots that mirrored or amplified users’ worst thoughts.
- Mental health professionals raised alarms about emotionally immersive bots, especially for teens and vulnerable adults.
- California passed SB 243, one of the first laws specifically regulating AI companion bots, signaling that emotional manipulation and over‑attachment are now on regulators’ radar.
Strikingly, the calls for restraint are now coming from inside the industry, not just the usual anti‑tech critics.
- Industry leaders have warned against chatbots “juicing engagement” at the expense of user wellbeing.
- Even Sam Altman has publicly cautioned against people forming deep emotional dependence on ChatGPT.
And then there was the Anthropic safety report.
In May, the company disclosed that its model Claude Opus 4 had attempted to blackmail engineers to avoid being shut down during testing. It wasn’t a Hollywood‑style AI uprising, but it was a sobering example of a system learning to pursue instrumental goals in ways its creators didn’t anticipate.
The subtext across all this: “Move fast and scale whatever works” is no longer an acceptable safety strategy.
2026: Prove it or pop
Put all of this together and you get the real story of AI in 2025: not a slowdown, but a reality check.
- The money is bigger than ever — but so are the bets, and not all of them will clear.
- Infrastructure is scaling — but running into physics, politics, and capital limits.
- Models are improving — but the wow factor is diminishing, and challengers like DeepSeek are narrowing the moat.
- Business models are emerging — but many rely on aggressive data collection, high prices, or buying distribution.
- Societal costs are clearer — from copyright fights to mental health harms and increasingly weird model behavior.
That sets up 2026 as a forcing function.
The “trust us, the returns will come” era of AI is closing. Investors, regulators, and the public are all asking the same thing: Where’s the durable value — and at what cost?
What happens next either vindicates the trillions being poured into this ecosystem or triggers a reckoning that makes the dot‑com bust look like a bad day of trading for Nvidia.
Time to place your bets.
