When OpenAI shut down its ChatGPT Atlas browser in August 2026, Perplexity was raising money at a $30B+ valuation to build Comet into the opposite bet — that the AI browser, not the chat window, is where the agent economy actually lives.
Nine months after launch, OpenAI shut down ChatGPT Atlas — its Chromium-based, agent-native browser — on August 9, 2026, folding whatever survived into the ChatGPT desktop app and a Chrome extension. The stated reason was almost disarmingly simple: the team decided the browser was "the wrong container" for the idea. Bookmarks and history didn't even migrate cleanly; users had to export an HTML file before the cutoff.
That same summer, Perplexity was doing the opposite. It raised roughly $200 million in June 2026 at a valuation near $20 billion, and by late August, Nvidia was reportedly in talks to invest at a valuation north of $30 billion — more than 50% above where Perplexity sat a year earlier. The money is going almost entirely toward one bet: that Comet, its own standalone AI browser, is the thing worth owning, not a feature bolted onto a chat app.
Two well-funded labs looked at the same question — should an AI company build its own browser? — and reached opposite conclusions within weeks of each other. That divergence is the cleanest way to understand what Perplexity actually is, because the company has never owned a frontier model. It has always had to make a bet about where in the stack the real value sits. Right now, that bet is: not the model, the surface.
Perplexity's product lineup has sprawled well past "AI search engine" over the past two years:
Revenue estimates for 2026 vary by source and month — Perplexity is private and doesn't publish official figures — but most trackers put annualized recurring revenue somewhere between $450 million and $750 million by mid-to-late 2026, up from roughly $230 million at the end of 2025. Whatever the exact number, the growth rate is the story: Pro/Max subscriptions reportedly still account for 60%+ of revenue, with enterprise contracts, Sonar API credits, and Comet Plus filling in the rest. A $400 million Snapchat integration deal and a $750 million three-year Microsoft Azure commitment, both announced in January 2026, round out a company that's spending like it expects the browser bet to pay off.
This is the detail that gets lost in valuation headlines: Perplexity doesn't train a frontier foundation model, and it isn't really trying to. It occupies the retrieval, ranking, and orchestration layer that sits on top of whichever models it chooses to call.
Perplexity has published details on the production pipeline behind Sonar, which the company says handles roughly 200 million queries a day. The design philosophy is explicitly about not choosing between two conventional approaches: rather than picking lexical (keyword) search or semantic (embedding) search, the first retrieval stage runs both against the index and merges the results into a hybrid candidate set. That stage optimizes for recall — cast a wide net — and hands off to increasingly selective ranking stages that optimize for precision. Any stage that risks becoming a latency bottleneck gets horizontally scaled rather than redesigned, which is a fairly unglamorous but telling engineering choice: at 200 million queries a day, boring and scalable beats clever and fragile.
The other structural decision is where citations live in the pipeline. Most RAG systems generate an answer and then attach sources as a post-hoc bibliography — a separate retrieval step feeding context into a generation step that has no obligation to track which sentence came from where. Perplexity's approach embeds citation tracking into the generation prompt and response logic itself, so each claim is tied to a source at generation time rather than reconstructed afterward. It's a meaningful difference in practice: post-hoc citation matching is why some AI answer engines cite a source that doesn't actually support the specific sentence next to it.
The Sonar model family reflects a build-vs-buy philosophy that spans both: Sonar and Sonar Pro are fine-tuned on open-weight bases (Perplexity's public documentation points to Llama derivatives) wrapped around the retrieval pipeline; Sonar Reasoning Pro is built on DeepSeek R1 for chain-of-thought analysis; Sonar Deep Research handles autonomous, multi-step research across hundreds of sources. Perplexity also released its own open-weight model, R1 1776, in February 2025 — a 671-billion-parameter post-trained version of DeepSeek-R1 with Chinese-government-aligned censorship removed, released under an MIT license on Hugging Face. It's a useful data point: Perplexity contributes to the open-model ecosystem, but as a downstream fine-tuner and distributor, not as a lab racing to the next GPT- or Gemini-class frontier checkpoint.
In February 2026, Perplexity dropped advertising entirely — no more sponsored answers against free-tier queries, a format it had only introduced in late 2024. According to reporting on the company's internal metrics, revenue jumped roughly 50% the following month. That's a counterintuitive result for anyone assuming ads are free money, but it lines up with a pattern seen elsewhere in consumer AI: sponsored placements in an answer engine create an incentive conflict that erodes the one thing Perplexity is actually selling — the credibility of "here's the cited, unbiased answer." Once that ad line disappeared, the entire business model consolidated around subscriptions, enterprise contracts, and API credits, which is a cleaner story to tell publishers, regulators, and enterprise buyers alike, even if it means walking away from a monetization channel every other consumer tech company treats as a default.
Comet's free worldwide rollout in October 2025 wasn't really about search share — Perplexity was never going to out-index Google. It was about owning the surface where an agent starts a task and, increasingly, where it finishes one: filling out a form, comparing prices across tabs, booking something, buying something. Whoever owns that surface owns the highest-intent moment in the agent economy, and that's the real prize behind the Nvidia talks.
| Perplexity (Comet) | Google (Gemini in Chrome / AI Mode) | OpenAI (post-Atlas) | You.com | |
|---|---|---|---|---|
| Owns a frontier foundation model | No — orchestrates third-party + fine-tuned open-weight models | Yes — Gemini | Yes — GPT family | No — orchestrates third-party models |
| Standalone AI-native browser | Yes — Comet, free since Oct. 2025 | No — Gemini embedded in existing Chrome | No — Atlas shut down Aug. 2026; features folded into ChatGPT app + Chrome extension | No — browser extension only |
| Publisher revenue share | Yes — Comet Plus, 80% of pooled subscription revenue to publishers | No formal program | No formal program | No formal program |
| Primary distribution moat | None inherited — has to win the browser from scratch | Chrome's existing ~3B+ user install base | ChatGPT's consumer install base and brand | Niche/enterprise search share |
| Reported 2026 valuation | ~$20-30B (in talks) | N/A (part of Alphabet) | ~$500B+ (reported) | Not publicly disclosed |
| Active copyright litigation | Yes — Britannica, News Corp/Dow Jones, NYT, CNN, others | Yes — separate suits | Yes — separate suits | Limited public litigation |
The table understates how different the underlying logic is. Google doesn't need a standalone browser bet because it already owns the browser — Gemini gets folded into Chrome's existing three-billion-plus install base. OpenAI concluded the opposite: rather than fight for browser share as a new entrant, it retreated to the surface it already dominates (the ChatGPT app itself) and shipped agentic browsing there instead, plus a Chrome extension for people who live in Google's browser anyway. Perplexity has neither Google's distribution nor OpenAI's brand recognition, so Comet is a much riskier bet by comparison — it has to win browser share from nothing, against Chrome's inertia, in a category where Microsoft (Edge, with Copilot), Google, and now a retreating OpenAI are all also playing.
Comet Plus is best understood as a business-model response to a legal fight it hasn't won. Perplexity faces active copyright suits from Dow Jones/News Corp (owner of the Wall Street Journal), Encyclopaedia Britannica and Merriam-Webster, the New York Times, CNN, and others, plus a separate suit from Amazon over disguised web scraping. None of these cases have reached trial or settlement as of September 2026. The Britannica/Merriam-Webster case, filed in the Southern District of New York in September 2025, is still in discovery: on September 2, 2026, a magistrate judge ordered Perplexity to hand over an additional snapshot of its RAG database and six months of additional user-activity-log data — over Perplexity's objection — while requiring the plaintiffs to help cover hosting costs for that data.
Comet Plus, launched with an initial $42.5 million payout pool, routes 80% of pooled Pro/Max/Comet Plus subscription revenue to participating publishers — an initial roster that includes CNN, the Washington Post, the Los Angeles Times, Le Monde, Le Figaro, Fortune, and Condé Nast's stable (The New Yorker, Wired, Vogue, Vanity Fair, Architectural Digest). It's a real payment mechanism, not just a press release, and it's a materially different structure than the licensing deals OpenAI and Google have separately negotiated with individual publishers. But it's also, transparently, litigation mitigation: the program only reached this form in January 2026, roughly eighteen months after Perplexity's first, smaller ad-revenue-sharing attempt and well into the lawsuits. A discovery order compelling more RAG and log data in September 2026 is a reminder that Comet Plus buys goodwill, not legal certainty — the underlying question of whether retrieval-and-synthesis over copyrighted text requires a license is still unresolved in court.
At a reported $23 billion valuation and $450 million-plus in estimated ARR, Perplexity was already trading at something like 50x revenue before the latest round of talks. A jump to $30 billion-plus pushes that multiple higher still, and it's worth being honest about what that means: the market isn't pricing Perplexity on its retrieval pipeline or its Sonar API business, both of which are genuinely good products with real developer traction. It's pricing the possibility that Comet becomes the default interface for agentic commerce — the place where an AI agent doesn't just answer your question but completes your purchase, and takes a cut. Nvidia's specific interest lines up with that: it's a strategic investor with its own interest in agentic-AI infrastructure demand, not a return-on-ARR calculation.
That's a bet on distribution and habit formation in a browser category that Google structurally dominates and that OpenAI just decided wasn't worth fighting for standalone. Perplexity is the one major AI company still willing to make that specific bet at scale — which makes it either the most interesting distribution play in consumer AI right now, or the company most exposed if the browser turns out to be, as OpenAI concluded, the wrong container after all.
The takeaway: Perplexity's valuation isn't a bet on model quality — it doesn't own a frontier model and isn't trying to. It's a bet that owning the browser, and paying publishers enough to keep the lawsuits from becoming existential, buys it the one thing Google and OpenAI can't easily hand it: the surface where the next generation of AI agents actually does things, not just answers questions.